A Deep Learning-Based Method and Apparatus for Wireless Radar Attitude Reconstruction

Abstract

This invention relates to the field of wireless radar equipment technology, and in particular to a method and apparatus for wireless radar attitude reconstruction based on deep learning. It acquires multiple frames of radar motion point clouds and wireless echo signals, uses echo intensity to guide the propagation of local coherent structures in the point clouds, achieving dynamic completion of sparse radar point clouds to obtain dynamically dense point cloud aggregates; constructs a smooth topological field and calculates structural loss through tensor decomposition to achieve structural alignment between the source and target point clouds, and identifies suspended point cloud regions based on attitude-invariant description; employs BEV (Bird's-eye View) enhanced coding and anchor frame detection to constrain coarse attitude view bounding boxes, and completes the inversion and backtracking of suspended point clouds through deep learning surface projection with cross-attention, obtaining attitude inversion projection points and suspended vectors; constructs a rigid skeleton and dynamic implicit space based on a seed attitude point cloud mesh, and uses suspended vectors to drive attitude reconstruction. This invention achieves dynamic reconstruction of wireless radar detection attitude, improving and optimizing attitude reconstruction accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of wireless radar equipment technology, and in particular to a wireless radar attitude reconstruction method and apparatus based on deep learning. Background Technology

[0002] Wireless radar, as an important means of environmental perception and target detection, is widely used in fields such as autonomous driving, unmanned systems, security monitoring, and industrial inspection. It acquires the distance, speed, and spatial distribution of targets by emitting electromagnetic waves and receiving target echo signals. With the increasing demands for dynamic perception accuracy in application scenarios, accurate reconstruction of wireless radar attitude has gradually become a research hotspot. In particular, dynamically updating the continuous attitude changes of wireless radar in complex dynamic environments has profound significance for improving the perception capabilities and decision-making levels of different detection systems.

[0003] Existing wireless radar attitude reconstruction methods mostly rely on single-frame or a small number of frame point cloud data, achieving attitude estimation through geometric matching, feature extraction, or simple temporal fusion. However, due to the inherent characteristics of radar information, such as low resolution, sparse data, and high noise, the reconstruction results are susceptible to interference from local missing data and outliers, making it difficult to guarantee the stability and accuracy of attitude estimation. When processing multi-frame continuous motion data, existing attitude reconstruction techniques often employ simple superposition or rigid registration strategies, failing to fully exploit the attitude motion information contained in the radar echo signal, resulting in insufficient information utilization. For "dangling points" or structurally discontinuous regions in radar motion point clouds caused by occlusion, viewpoint changes, etc., there is a lack of effective identification and compensation mechanisms, which easily leads to mismatches, thereby reducing the accuracy of attitude reconstruction. In addition, although some deep learning-based reconstruction methods have improved feature representation, they often neglect the collaborative modeling of global structural constraints and local geometric consistency, making it difficult to balance global topology and detailed representation in the reconstruction results, resulting in poor robustness and an inability to achieve high-precision attitude reconstruction in complex motion or rigidly changing scenarios. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a method and apparatus for wireless radar attitude reconstruction based on deep learning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a wireless radar attitude reconstruction method based on deep learning, comprising the following steps: S01: Acquire multiple frames of radar motion point cloud and wireless echo signal from the previous scan timestamp to the current scan timestamp of the wireless radar. Guide the local coherent structure propagation of the radar motion point cloud by the echo intensity of the attitude motion information contained in the wireless echo signal, complete the sparse point cloud motion, and obtain the dynamic dense point cloud set of the wireless radar motion. S02: Construct the topological structure of the smooth dynamic dense point cloud set. Reconstruct the smooth topological field of the target. Calculate the structural loss between the target smooth topological field and the source smooth topological field through tensor decomposition. Minimize the structural loss and align the dynamic point cloud structure of the motion. Under the premise of alignment, analyze the point cloud distribution with attitude change in the point cloud structure to determine the suspended point cloud region. S03: Apply the offset vector features of the moving point cloud in the source reconstruction smooth topology field using BEV bird's-eye view plane enhancement coding source and perform deep learning detection of the offset point cloud anchor box to obtain the view bounding box of the coarse motion attitude of the wireless radar. Interact the suspended point cloud area onto the surface projection backtracking of the source reconstruction smooth topology field with cross attention. Use the view bounding box as a coarse attitude constraint to drive the relative displacement of the surface projection to obtain the attitude inversion projection points and their suspended vectors of the suspended point cloud area. S04: Obtain the seed attitude point cloud mesh of the previous scan timestamp of the wireless radar, construct a rigid skeleton and its dynamic transformation implicit space based on the motion degrees of freedom of the wireless radar and the attitude inversion projection points, dynamically transform the rigid skeleton based on the suspended vector and inject it into the dynamic transformation implicit space, reconstruct the attitude rigid transformation of the suspended point cloud starting from the seed attitude point cloud mesh, and output the current reconstructed attitude information of the wireless radar.

[0006] Furthermore, S01 specifically includes the following steps: The system acquires multiple frames of radar motion point cloud continuously collected from the previous scan timestamp to the current scan timestamp during the time segment of the target scanning scene detected by the wireless radar, as well as the wireless echo signal corresponding to each frame of radar motion point cloud. At the same time, it acquires the predetermined sensing clock when the wireless radar outputs the motion echo signal. Based on the predetermined sensing clock, a motion sensing coordinate system of the wireless radar at the current scanning timestamp is constructed. A short-time Fourier transform algorithm is introduced to perform spatial frequency domain feature analysis on the wireless echo signal of each frame to obtain the signal transmit / receive vector, echo propagation time, frequency offset and echo array phase difference. Based on echo propagation time, frequency offset and echo array phase difference, the spatial transformation between radar moving point clouds in each frame is estimated. During the estimation process, the echo energy distribution contributing to the physical spatial location of each wireless echo signal is weighted according to the spatial transformation attribute to obtain the echo intensity guidance weight value. A weighted radial basis function field for spatial point cloud motion is constructed using PROE model software and multiple quadratic radial basis kernels. Based on echo intensity-guided weights, each wireless echo signal is introduced into the weighted radial basis function field to obtain a continuous radial basis coefficient vector for motion continuity. The coherent radial basis coefficient vector is used to interpolate and propagate the solution to the local homogeneous neighborhood of each radar motion point cloud in the structural direction, generating the sparse completion attributes and the probability distribution of the completion points for each frame of radar motion point cloud; wherein, the sparse completion attributes include the motion coordinates, spatial distance and structural representation level of the completion point cloud. Based on the sparse completion attribute and the probability distribution of the completion points, the radar motion point cloud of each frame is transformed and fixed on the motion sensing coordinate system for temporal fusion and registration, filling in the sparse and empty point cloud motion narrative, and obtaining the dynamic dense point cloud set of the wireless radar motion.

[0007] Furthermore, S02 specifically includes the following steps: By estimating the density characteristics of a dynamic dense point cloud, the elastic energy function of the radius neighborhood topology is defined based on the spatiotemporal Euclidean distance and dynamic integral iteration is performed based on it to construct a topological density attraction domain. The backbone topology is established by the topological connectivity partitions of the topological density attraction domain and the surrogate saddle points. The dynamic dense point cloud is smoothly distributed to the backbone topology to form a smooth topological field of point cloud motion of the dynamic dense point cloud, which is defined as the target reconstructed smooth topological field. Obtain the smoothed topological field of the point cloud motion at the previous scan timestamp of the wireless radar, and define it as the source reconstructed smoothed topological field; Tensor decomposition algorithm is introduced to decompose and calculate the source reconstructed smooth topological field and the target reconstructed smooth topological field respectively, to obtain the multidimensional structure loss tensor of the source reconstructed smooth topological field, which is labeled as the first structure loss tensor; and to obtain the multidimensional structure loss tensor of the target reconstructed smooth topological field, which is labeled as the second structure loss tensor. Calculate the structural loss gradient between the first structural loss tensor and the second structural loss tensor. Based on the structural loss gradient, preset the effective cost matrix of the motion structure of the target reconstructed smooth topology field relative to the source reconstructed smooth topology field. Align and match the point cloud structure of the source reconstructed smooth topology field to the target reconstructed smooth topology field. During the alignment matching process, the optimal matching transportation relationship of the point cloud structure is updated based on the effective cost matrix, and the structural hash degree of the target reconstructed smooth topology field relative to the source reconstructed smooth topology field is determined. The weighted covariance matrix between point cloud structures is calculated based on the criterion of minimizing structural scattering. A local point cloud distribution reference frame is constructed based on the weighted covariance matrix. Under the local point cloud distribution reference frame, the point cloud space is collaboratively divided into N sub-point cloud distribution domains. The angle between the normal vectors of different point clouds in the source reconstructed smooth topological field and the target reconstructed smooth topological field within each sub-point cloud distribution domain is calculated by histogram quantization and described and extracted, generating a suspended isosurface with attitude-invariant description in each sub-point cloud distribution domain. If the area of ​​the suspended isosurface is greater than the preset area value, then the sub-point cloud distribution domain is marked as a suspended point cloud region with attitude changes.

[0008] Furthermore, by estimating the density characteristics of the dynamically dense point cloud, defining the elastic energy function of the radius neighborhood topology based on the spatiotemporal Euclidean distance, and performing dynamic integration iteration based on it, a topological density attraction domain is constructed. A backbone topological graph is established through the topological connectivity partitions of the topological density attraction domain and surrogate saddle points. The dynamically dense point cloud is then smoothly distributed onto the backbone topological graph, forming a smooth topological field for the motion of the dynamically dense point cloud. Specifically, this includes the following steps: By constructing a spatial scalar potential field using PROE model software, and introducing a kernel density algorithm to estimate the continuous density and potential energy troughs of the input spatial scalar potential field of the dynamic dense point cloud, a density gradient vector field is obtained. Construct a radius neighborhood topology for a dynamic dense point cloud, obtain the time-varying spatiotemporal Euclidean distance between every two adjacent dynamic dense point clouds, and define the elastic energy function of the radius neighborhood topology based on the spatiotemporal Euclidean distance. The density gradient vector field is used to perform dynamic integral iteration on each dynamic dense point cloud in the dynamic dense point cloud set based on the elastic energy function, and the current local density value is output by dynamic maintenance. If the local density value has already converged to the center of the maximum local density value, then immediately stop the iterative operation of the dynamic integral, generate the structural convergence path of each dynamic dense point cloud corresponding to the topological center and the modal nodes along the path, and couple the structural convergence path of each dynamic dense point cloud to the modal nodes to construct the topological density attraction domain. The topological connectivity partitions and topological surrogate saddle points that change with time by topological density attraction domains are obtained. The main topological graph is jointly constructed with the topological surrogate saddle points as endpoints and the topological connectivity partitions as connecting edges. A continuous motion topology description code is compiled based on spatiotemporal Euclidean distance, and each dynamic dense point cloud is assigned and mapped onto the backbone topology map. The motion topology description code is then attached to the point cloud and smoothly distributed, ultimately forming a smooth motion topology field of the point cloud set.

[0009] Furthermore, S03 specifically includes the following steps: Introducing the BEV bird's-eye view plane, the source reconstructed smooth topological field is divided into k topological anchor pillars on the BEV bird's-eye view plane. The offset vectors of different moving point clouds within each topological anchor pillar are feature-enhanced encoded using a point cloud network, and the BEV pillar features of several candidate offset point clouds are output. Each BEV support feature is back-mapped to the predetermined position of the corresponding candidate offset point cloud on the BEV bird's-eye view plane to obtain the BEV pseudo offset support. The position branch and orientation branch of the BEV pseudo offset support are detected by CNN convolutional neural network and anchor-based anchor box detection method, and finally the view bounding box of the coarse motion attitude of the wireless radar is obtained. Each suspended point cloud in each of the suspended point cloud regions is interactively projected onto the motion structure surface represented by the source reconstructed smooth topological field. During the interactive projection process, the cross-attention information of the surface projection is simultaneously extracted from the source reconstructed smooth topological field to obtain the back-back projection representation of each suspended point cloud and its cross-attention weight. The back-projection representation includes the back-projection query vector, source key vector, and value vector of the suspended point cloud; The local surface features of each suspended point cloud region are aligned and aggregated based on the back-back projection representation to obtain the implicit attention surface. The relative displacement vector of the suspended point cloud is predicted using the implicit attention surface based on the cross attention weight, and the surface projection estimation result is generated. By introducing the source reconstruction smooth topological field as a geometric framework constraint using the view bounding box, the surface projection estimation results are projected along the coarse attitude gradient direction of the source reconstruction smooth topological field, thus obtaining the attitude inversion projection points and their dangling vectors of the dangling point cloud region.

[0010] Furthermore, S04 specifically includes the following steps: Obtain the seed attitude point cloud mesh of the wireless radar when reconstructing the smooth topology field from the source generated at the previous scan timestamp; Obtain the degrees of freedom of motion of the wireless radar itself or the detection carrier to which the wireless radar is attached, and construct a rigid skeleton of the wireless radar attitude based on the degrees of freedom of motion. Based on the attitude inversion projection points, the transformation implicit function of the rigid skeleton is defined. The influence of the attitude skeleton of different suspended point clouds in the suspended point cloud region is propagated and combined from the seed attitude point cloud mesh using the transformation implicit function to construct the dynamic transformation implicit space of the rigid skeleton. Based on the suspended vector, a transformation animation operation based on the degree of freedom of motion is applied to the rigid skeleton to obtain the motion transformation matrix of the rigid skeleton. The inverse mapping algorithm is introduced to inject the motion transformation matrix into the dynamic transformation implicit space to obtain the pre-transformation field value of each suspended point cloud. Based on the pre-transformation field value, each suspended point cloud is projected and propagated from the seed attitude point cloud mesh along the normal trajectory direction of the attitude inversion projection point to its corresponding suspended isosurface, thus obtaining the dynamic propagation matrix of attitude rigid transformation. Based on the dynamic propagation matrix, the point cloud positions of the seed attitude point cloud mesh are iteratively maintained and reconstructed through motion rigid transformation, and finally the current reconstructed attitude information of the wireless radar is output.

[0011] A second aspect of the present invention provides a deep learning-based wireless radar attitude reconstruction device, the device comprising: a memory, a processor, and a communication interface, wherein the memory includes a deep learning-based wireless radar attitude reconstruction method program, and the communication interface is used for data connection communication between the memory and the processor, wherein when the wireless radar attitude reconstruction method program is executed by the processor, it implements any of the steps of the wireless radar attitude reconstruction method described in the present invention.

[0012] This invention addresses the technical deficiencies in the prior art, and its beneficial technical effects are as follows: By guiding the propagation of the local coherent structure of the radar moving point cloud through echo signals, dynamic completion of sparse radar information is achieved, constructing a high-density, more continuous dynamic point cloud set, effectively alleviating the sparsity and discreteness of traditional radar point clouds. By constructing a smooth topological field and combining it with tensor decomposition for structural loss constraints, precise alignment of the source and target point clouds at the topological level is achieved. On this basis, attitude-invariant description is introduced to analyze the point cloud distribution, which can accurately identify hard-to-detect suspended point cloud regions, avoid attitude estimation deviations caused by local missing or occluded areas, and improve the structural perception capability and reconstruction accuracy in complex scenes. Using BEV bird's-eye view plane augmented coding and anchor box detection mechanisms, the offset features of the moving point cloud are efficiently expressed. Combined with a deep learning strategy of surface projection backtracking with cross-attention, attitude inversion of suspended point clouds is achieved under the coarse attitude constraints of the view bounding box. This effectively integrates global view constraints and local geometric details, making the attitude estimation not only globally consistent but also finely characterizing local structural changes, significantly improving the accuracy and stability of attitude inversion. By constructing a rigid skeleton and its implicit dynamic transformation space, and using suspended vectors to drive skeleton transformation, continuous rigid transformation reconstruction from seed attitude point cloud mesh to the current attitude is achieved. This enables high-fidelity dynamic modeling and updating of the wireless radar while ensuring motion rationality, effectively avoiding reconstruction errors caused by insufficient rigidity assumptions or non-rigid interference in traditional methods. This invention significantly improves the accuracy and robustness of wireless radar attitude reconstruction and enhances its adaptability to complex dynamic scenes and sparse radar information conditions. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0014] Figure 1A flowchart of the first method for wireless radar attitude reconstruction based on deep learning is shown. Figure 2 A flowchart of the second method for wireless radar attitude reconstruction based on deep learning is shown; Figure 3 A structural diagram of a deep learning-based wireless radar attitude reconstruction device is shown. Detailed Implementation

[0015] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0016] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0017] The first aspect of this invention provides a wireless radar attitude reconstruction method based on deep learning, such as... Figure 1 As shown, it includes the following steps: S01: Acquire multiple frames of radar motion point cloud and wireless echo signal from the previous scan timestamp to the current scan timestamp of the wireless radar. Guide the local coherent structure propagation of the radar motion point cloud by the echo intensity of the attitude motion information contained in the wireless echo signal, complete the sparse point cloud motion, and obtain the dynamic dense point cloud set of the wireless radar motion. S02: Construct the topological structure of the smooth dynamic dense point cloud set. Reconstruct the smooth topological field of the target. Calculate the structural loss between the target smooth topological field and the source smooth topological field through tensor decomposition. Minimize the structural loss and align the dynamic point cloud structure of the motion. Under the premise of alignment, analyze the point cloud distribution with attitude change in the point cloud structure to determine the suspended point cloud region. S03: Apply the offset vector features of the moving point cloud in the source reconstruction smooth topology field using BEV bird's-eye view plane enhancement coding source and perform deep learning detection of the offset point cloud anchor box to obtain the view bounding box of the coarse motion attitude of the wireless radar. Interact the suspended point cloud area onto the surface projection backtracking of the source reconstruction smooth topology field with cross attention. Use the view bounding box as a coarse attitude constraint to drive the relative displacement of the surface projection to obtain the attitude inversion projection points and their suspended vectors of the suspended point cloud area. S04: Obtain the seed attitude point cloud mesh of the previous scan timestamp of the wireless radar, construct a rigid skeleton and its dynamic transformation implicit space based on the motion degrees of freedom of the wireless radar and the attitude inversion projection points, dynamically transform the rigid skeleton based on the suspended vector and inject it into the dynamic transformation implicit space, reconstruct the attitude rigid transformation of the suspended point cloud starting from the seed attitude point cloud mesh, and output the current reconstructed attitude information of the wireless radar.

[0018] Furthermore, S01 specifically includes the following steps: The system acquires multiple frames of radar motion point cloud continuously collected from the previous scan timestamp to the current scan timestamp during the time segment of the target scanning scene detected by the wireless radar, as well as the wireless echo signal corresponding to each frame of radar motion point cloud. At the same time, it acquires the predetermined sensing clock when the wireless radar outputs the motion echo signal. Based on the predetermined sensing clock, a motion sensing coordinate system of the wireless radar at the current scanning timestamp is constructed. A short-time Fourier transform algorithm is introduced to perform spatial frequency domain feature analysis on the wireless echo signal of each frame to obtain the echo propagation time, frequency offset and echo array phase difference. Based on echo propagation time, frequency offset and echo array phase difference, the spatial transformation between radar moving point clouds in each frame is estimated. During the estimation process, the echo energy distribution contributing to the physical spatial location of each wireless echo signal is weighted according to the spatial transformation attribute to obtain the echo intensity guidance weight value. A weighted radial basis function field for spatial point cloud motion is constructed using PROE model software and multiple quadratic radial basis kernels. Based on echo intensity-guided weights, each wireless echo signal is introduced into the weighted radial basis function field to obtain a continuous radial basis coefficient vector for motion continuity. The coherent radial basis coefficient vector is used to interpolate and propagate the solution to the local homogeneous neighborhood of each radar motion point cloud in the structural direction, generating the sparse completion attributes and the probability distribution of the completion points for each frame of radar motion point cloud; wherein, the sparse completion attributes include the motion coordinates, spatial distance and structural representation level of the completion point cloud. Based on the sparse completion attribute and the probability distribution of the completion points, the radar motion point cloud of each frame is transformed and fixed on the motion sensing coordinate system for temporal fusion and registration, filling in the sparse and empty point cloud motion narrative, and obtaining the dynamic dense point cloud set of the wireless radar motion.

[0019] It should be noted that wireless radar typically relies on the scattering of electromagnetic waves by the target. Most surfaces will slowly scatter or absorb the waves, such as due to energy dispersion or factors like clothing and human tissue. Only radar detection locations that meet strong reflection conditions will produce significant echoes. Secondly, the resolution of radar is limited by bandwidth and array size, resulting in typically low angular resolution. When scanning and detecting targets at long distances, multiple real-world points are easily merged into a single observation point. In complex scenes, obstructions can make some targets invisible, causing strong sparsity in radar information. This leads to low-density, incomplete, discontinuous, and insufficient representation of the scene point cloud, making it impossible to accurately reconstruct the attitude and motion of the wireless radar itself. To address this, this method acquires continuously sampled radar motion point clouds and their transmitted and received wireless echo signals. The radar motion point cloud represents the observation perspective information of the scene scanned by the wireless radar. The attitude information contained in the wireless echo signal mainly includes: direction (signal transmission and reception vectors), distance (echo propagation time), velocity (frequency offset), and angle (array phase difference). The time reference for each frame of point cloud is unified based on the sensor timestamp of the wireless radar, i.e., a predetermined sensing clock, to avoid spatial mismatch caused by time misalignment and compensate for radar sensor motion. The relative pose transformation between each frame of moving radar point cloud is estimated using information such as the direction, distance, velocity, and angle of the echo signal, exploring the spatial motion of the wireless radar sensor at different time nodes. This provides a reliable temporal accumulation premise and spatial transformation basis for spatial point cloud registration. Spatial transformation attributes include translation, scaling, rotation, affine transformation, projection transformation, and similarity transformation. By mapping echo intensity to weights, a weighted echo intensity-guided weight value is constructed, giving different echo signals a physically perceptible propagation guide for describing the spatial pose change of the wireless radar. This allows the radar signal to guide the point cloud for physical-level transformation supplementation. The radar point cloud completion interpolation breaks away from the pure geometric propagation of traditional techniques, increasing the contribution of high-quality echoes to point cloud completion interpolation, significantly improving physical reliability, and ensuring the sparse completion gain of radar information.

[0020] It should be noted that the multiple quadratic radial basis kernels can define the influence of spatial distance on signal guidance propagation. The constructed weighted radial basis function field allows echo information to be embedded within the radial basis functions, determining the smoothness and global characteristics of signal-guided spatial point cloud propagation. Through a coherent radial basis coefficient vector, linear propagation dependent on the timing of the echo signal is achieved, rather than random, blind point filling. By solving for the propagation of each radar motion point cloud based on the radar echo signal, its sparse regions are gradually "filled," with information spreading from known points, thus filling in the missing content in the point cloud caused by the sparse radar information. This method enables the global continuous propagation of spatial point clouds guided by the echo signal of the wireless radar, filling in the missing point cloud during the radar attitude transformation process, effectively compensating for the sparse characteristics of radar information, and ensuring the accuracy and realism of attitude reconstruction.

[0021] Furthermore, S02 specifically includes the following steps: By estimating the density characteristics of a dynamic dense point cloud, the elastic energy function of the radius neighborhood topology is defined based on the spatiotemporal Euclidean distance and dynamic integral iteration is performed based on it to construct a topological density attraction domain. The backbone topology is established by the topological connectivity partitions of the topological density attraction domain and the surrogate saddle points. The dynamic dense point cloud is smoothly distributed to the backbone topology to form a smooth topological field of point cloud motion of the dynamic dense point cloud, which is defined as the target reconstructed smooth topological field. Obtain the smoothed topological field of the point cloud motion at the previous scan timestamp of the wireless radar, and define it as the source reconstructed smoothed topological field; Tensor decomposition algorithm is introduced to decompose and calculate the source reconstructed smooth topological field and the target reconstructed smooth topological field respectively, to obtain the multidimensional structure loss tensor of the source reconstructed smooth topological field, which is labeled as the first structure loss tensor; and to obtain the multidimensional structure loss tensor of the target reconstructed smooth topological field, which is labeled as the second structure loss tensor. Calculate the structural loss gradient between the first structural loss tensor and the second structural loss tensor. Based on the structural loss gradient, preset the effective cost matrix of the motion structure of the target reconstructed smooth topology field relative to the source reconstructed smooth topology field. Align and match the point cloud structure of the source reconstructed smooth topology field to the target reconstructed smooth topology field. During the alignment matching process, the optimal matching transportation relationship of the point cloud structure is updated based on the effective cost matrix, and the structural hash degree of the target reconstructed smooth topology field relative to the source reconstructed smooth topology field is determined. The weighted covariance matrix between point cloud structures is calculated based on the criterion of minimizing structural scattering. A local point cloud distribution reference frame is constructed based on the weighted covariance matrix. Under the local point cloud distribution reference frame, the point cloud space is collaboratively divided into N sub-point cloud distribution domains. The angle between the normal vectors of different point clouds in the source reconstructed smooth topological field and the target reconstructed smooth topological field within each sub-point cloud distribution domain is calculated by histogram quantization and described and extracted, generating a suspended isosurface with attitude-invariant description in each sub-point cloud distribution domain. If the area of ​​the suspended isosurface is greater than the preset area value, then the sub-point cloud distribution domain is marked as a suspended point cloud region with attitude changes.

[0022] It should be noted that, since observations in wireless radar sensing typically rely on the scattering characteristics of the target, attitude changes not only correspond to rigid body transformations but also alter the set and intensity distribution of observable scattering centers through angle dependence and occlusion effects. This results in the same object exhibiting significantly different or even nearly unrelated point cloud structures under different attitudes, with completely different point cloud distributions. In the absence of rotational invariance or isovariability modeling, the object is often misclassified as a different category (different object) under the condition of distribution differences caused by attitude changes. This leads to unstable attitude estimation, decreased cross-attitude generalization ability, and reconstruction bias or even failure, resulting in attitude reconstruction not being attached to the real scene structure, and the reconstruction results being inaccurate. To address this, our method first decomposes the structural loss tensors of the target reconstructed smooth topological field and the source reconstructed smooth topological field using tensor decomposition. The structural loss tensor defines the structural consistency of the two across spatial alignment, which describes the structural mapping relationship of temporal point cloud pairs in the spatial scene. Furthermore, the calculated structural loss gradient ensures that the spatial matching result of the wireless radar attitude from the previous scan timestamp to the current scan timestamp is only constrained by the global geometric relationship, ignoring the influence of the medium of rotation or translation, and eliminating the sensitivity to attitude-induced changes (i.e., maintaining the point cloud structure distribution with unchanged attitude). Furthermore, due to the extremely sparse nature, strong noise, and discrete properties of wireless radar detection, point cloud topology is prone to fragmentation, local noise is amplified, and severe oscillations and instabilities in the structural field are generated. This results in a non-smooth distribution of the point cloud structure characterizing radar motion, characterized by neighborhood discontinuities, missing segments, and non-geometric consistency. It is difficult to accurately describe the essential structure of the scene and can easily exacerbate or mislead observation errors in attitude reconstruction. Therefore, this method smooths the potentially sparse, unstable, and discrete wireless radar observations (dynamically dense point clouds) to reconstruct a smooth topological field for the target, forming a continuous, stable, and optimizable point cloud distribution topological representation. This improves the solvability and robustness of attitude reconstruction and avoids the derivation of reconstruction observation errors.

[0023] It should be noted that by constructing the effective cost matrix of the current relative structural loss, the effective cost matrix quantifies the cost of mapping the difference in point cloud distribution structure to the point pair matching cost under the current matching condition, under the premise of attitude invariance. It details the allocation of the degree of unreasonable matching between each point cloud pair, guiding the direction of advanced exploration under attitude invariance. After fixing the effective cost of the temporal point cloud distribution of radar motion, the matching relationship between moving point cloud pairs is updated. This avoids directly aligning the moving point clouds themselves, but rather aligns the topological structure of the point cloud distribution, i.e., the consistent spatiotemporal distance relationship between the point cloud at the previous timestamp and the point cloud at the current timestamp. This strengthens structurally consistent matching and suppresses structurally inconsistent matching, forming an iterative closed-loop mechanism of structural solidification. Thus, during attitude transformation inference, the perspective shifts to the abnormal locality of point cloud distribution under the premise of fixed structure, improving the focus on the point cloud distribution structure. When minimizing structural hashing to maintain topological alignment, the distribution of moving point clouds under different attitude changes of the same wireless radar maintains global consistency to the maximum extent. However, if the point cloud distribution is completely different even under structural alignment, it indicates that the point cloud at that location is a false point in the scene after the wireless radar's attitude change, which is crucial for attitude reconstruction. Therefore, by using distance weighting of the point cloud structure neighborhood through covariance statistics, the structural contribution of the point cloud near the center is highlighted, and a local point cloud distribution reference frame is constructed to enhance the stability of local principal direction estimation. Furthermore, feature decomposition of the point cloud space of the target reconstruction smooth topological field and the source reconstruction smooth topological field is performed to extract principal directions, obtaining an orthogonal basis describing the local geometric distribution, characterizing the local distribution trend of the moving point cloud, and recovering stable geometric structure orientation information from the disordered moving point cloud. Finally, the angle between the normal of the neighborhood point cloud and the normal of the key point cloud is calculated to characterize the distribution change characteristics of the local point cloud surface, forming an attitude-invariant descriptor, i.e., a suspended isosurface. If the area of ​​the suspended isosurface is greater than a preset area value, it indicates that the moving point cloud in this distribution area represents the viewpoint structure of the wireless radar, which is not in the real scene. This suggests that it is false information caused by attitude transformation, providing quantifiable clues for attitude reconstruction. This method can locate the abnormal (suspended) point cloud distribution caused by attitude transformation under the premise of consistent point cloud structure alignment. This clarifies the attitude motion transformation path and logic of the wireless radar, reduces the sensitivity to attitude dependence, and improves the accuracy of attitude reconstruction analysis.

[0024] Furthermore, by estimating the density characteristics of the dynamically dense point cloud, defining the elastic energy function of the radius neighborhood topology based on the spatiotemporal Euclidean distance, and performing dynamic integration iteration based on it, a topological density attraction domain is constructed. A backbone topological graph is established through the topological connectivity partitions of the topological density attraction domain and surrogate saddle points. The dynamically dense point cloud is then smoothly distributed onto the backbone topological graph, forming a smooth topological field for the motion of the dynamically dense point cloud. Specifically, this includes the following steps: By constructing a spatial scalar potential field using PROE model software, and introducing a kernel density algorithm to estimate the continuous density and potential energy troughs of the input spatial scalar potential field of the dynamic dense point cloud, a density gradient vector field is obtained. Construct a radius neighborhood topology for a dynamic dense point cloud, obtain the time-varying spatiotemporal Euclidean distance between every two adjacent dynamic dense point clouds, and define the elastic energy function of the radius neighborhood topology based on the spatiotemporal Euclidean distance. The density gradient vector field is used to perform dynamic integral iteration on each dynamic dense point cloud in the dynamic dense point cloud set based on the elastic energy function, and the current local density value is output by dynamic maintenance. If the local density value has already converged to the center of the maximum local density value, then immediately stop the iterative operation of the dynamic integral, generate the structural convergence path of each dynamic dense point cloud corresponding to the topological center and the modal nodes along the path, and couple the structural convergence path of each dynamic dense point cloud to the modal nodes to construct the topological density attraction domain. The topological connectivity partitions and topological surrogate saddle points that change with time by topological density attraction domains are obtained. The main topological graph is jointly constructed with the topological surrogate saddle points as endpoints and the topological connectivity partitions as connecting edges. A continuous motion topology description code is compiled based on spatiotemporal Euclidean distance, and each dynamic dense point cloud is assigned and mapped onto the backbone topology map. The motion topology description code is then attached to the point cloud and smoothly distributed, ultimately forming a smooth motion topology field of the point cloud set.

[0025] It should be noted that the smoothing process for dynamically dense point clouds involves kernel density estimation in a spatial scalar potential field to clarify the density profile of the dynamically dense point cloud. This transforms the discontinuous and sparse dynamically dense point cloud into a smooth density distribution, i.e., a density gradient vector field. Dense regions of the point cloud represent potential scene motion structures, while sparse regions may represent void factors or boundary elements. The radius neighborhood topology serves as the topological framework for elastic energy. By defining the elastic energy function of the radius neighborhood topology using the spatiotemporal Euclidean distance between adjacent dynamically dense point clouds, the discrete dense point cloud is transformed into a structural system capable of transmitting energy, determining the smooth energy propagation path of topological information. In short, the core concept of this step is to treat a dynamic dense point cloud as an elastic body, with adjacent point clouds acting like "springs." Deviating from the original structure generates energy. Therefore, by constructing and minimizing its energy function, the sparsity suppression cost and continuity structure constraint of the dynamic dense point cloud are characterized, enabling the topological result of the point cloud distribution to have the ability to maintain structure rather than the traditional pure smoothness. At the same time, dynamic iteration based on the elastic energy function realizes the coupling of point cloud information from local to global, anchors the core transition carrier of topology propagation, and prevents oversmoothing and structural distortion while maintaining the complete narrative of point cloud motion, thus improving the smoothness of the dynamic dense point cloud with reliable topological accuracy.

[0026] It should be noted that the current local density value quantifies the offset vector of the current point cloud pointing towards the direction of maximum local density growth, defining a structural guide for unsupervised movement direction for each dynamically dense point cloud, thus optimizing topological accuracy. If the current local density value has converged and gathered to the center of the maximum local density value, it indicates that the dynamically dense point cloud has automatically gathered to the modal center, forming a significant structural contraction phenomenon. Here, the modal nodes correspond to the structural core in the actual dynamic point cloud distribution, serving as anchor points for the topological structure. Based on the path and modality to which each point ultimately converges, the dynamic point cloud space is discretized into topological density attraction domains, forming a transition trend in the regional topological representation. Topological connectivity partitions and topological surrogate saddle points provide topological understanding information, which can improve the richness of the structural representation of topological smoothness in dynamically dense point clouds.

[0027] Furthermore, S03, as Figure 2 As shown, the specific steps include: Introducing the BEV bird's-eye view plane, the source reconstructed smooth topological field is divided into k topological anchor pillars on the BEV bird's-eye view plane. The offset vectors of different moving point clouds within each topological anchor pillar are feature-enhanced encoded using a point cloud network, and the BEV pillar features of several candidate offset point clouds are output. Each BEV support feature is back-mapped to the predetermined position of the corresponding candidate offset point cloud on the BEV bird's-eye view plane to obtain the BEV pseudo offset support. The position branch and orientation branch of the BEV pseudo offset support are detected by CNN convolutional neural network and anchor-based anchor box detection method, and finally the view bounding box of the coarse motion attitude of the wireless radar is obtained. Each suspended point cloud in each of the suspended point cloud regions is interactively projected onto the motion structure surface represented by the source reconstructed smooth topological field. During the interactive projection process, the cross-attention information of the surface projection is simultaneously extracted from the source reconstructed smooth topological field to obtain the back-back projection representation of each suspended point cloud and its cross-attention weight. The back-projection representation includes the back-projection query vector, source key vector, and value vector of the suspended point cloud; The local surface features of each suspended point cloud region are aligned and aggregated based on the back-back projection representation to obtain the implicit attention surface. The relative displacement vector of the suspended point cloud is predicted using the implicit attention surface based on the cross attention weight, and the surface projection estimation result is generated. By introducing the source reconstruction smooth topological field as a geometric framework constraint using the view bounding box, the surface projection estimation results are projected along the coarse attitude gradient direction of the source reconstruction smooth topological field, thus obtaining the attitude inversion projection points and their dangling vectors of the dangling point cloud region.

[0028] It should be noted that after identifying the suspended point cloud regions, deep learning is used to infer the attitude transformation of these regions. However, existing deep learning methods typically lack explicit correspondences for the global transformation features of suspended point clouds, and can only rely on statistical features for attitude guessing. Furthermore, traditional deep learning models default to observing complete objects, and the forced matching using global transformation features leads to severe deviations in attitude evolution. Therefore, this method projects the cross-attention of suspended point clouds onto the surface of the query point cloud (source reconstructed smooth topological field) representation, outputting a back-back projection representation that realistically infers the suspended point cloud inversion from the motion point cloud distribution of the previous scan timestamp after the wireless radar attitude transformation. These back-back projection representations are globally learnable soft neighborhood candidates for the starting point cloud distribution of the source reconstructed smooth topological field with respect to the suspended point clouds. Compared with traditional deep learning attitude reconstruction techniques, this method eliminates unstable neighborhood inference components and accurately captures the non-local structure caused by attitude transformation. The implicit attention surface describes the local surface features of different suspended point clouds that are of interest to the query point cloud (source reconstructed smooth topological field) in terms of attitude perception. These features include curvature trends, surface orientation, and neighborhood distribution, anchoring the precise location of the reverse suspended backtracking trajectory and its landing point indication. Based on cross-attention and aggregation features, the relative displacement vector is predicted, thereby achieving attitude inversion mapping from the suspended point cloud to the source reconstructed smooth topological field representation. This restores the continuous surface of the attitude transformation trajectory and quickly locks the projection point and suspended vector.

[0029] It should be noted that traditional deep learning is sensitive to the initial pose and is prone to convergence to incorrect pose paths. This leads to mismatches between target and background points, causing topological features to be mixed with irrelevant points and resulting in complete distortion. This exposes the limitation of blind search across the entire space of the continuous surface during pose transformation. To address this, our method uses the BEV (Bird's-Eye View) plane to perform fine-grained partitioning of the source reconstructed smooth topological field. BEV has excellent characteristics in capturing object motion and has the potential to supplement point-by-point representations, transforming the disordered point cloud into structured spatial units, namely topological anchor pillars. Each topological anchor pillar fixes the point cloud distribution set of different local regions, preserving the target scanning scene and the initial horizontal spatial resolution of the wireless radar. Furthermore, the geometric offset relationships of the moving point clouds within the topological anchor pillars are encoded into high-dimensional features, including offsets relative to the center of the topological anchor pillar and offsets relative to the centroid of the point cloud. This reduces the sensitivity of deep learning to the point cloud order when inferring wireless radar pose transformations, effectively eliminating incorrect convergence of the pose path. Next, the BEV support features are placed back into the corresponding grid positions on the BEV bird's-eye view plane. The position and orientation branches of the BEV pseudo-offset support are then detected by regression using a CNN convolutional neural network and an anchor-based detection method. The output is a 3D bounding box parameter, which includes the detected target's position, orientation, and size. This forms a coarse attitude framework from the main viewpoint of the wireless radar, which imposes interpretable initial geometric constraints on the driving evolution of surface projection estimation. This prevents phenomena such as projection drift and infinite noise amplification during deep learning, and has high generalization ability, improving the accuracy of attitude evolution and transformation source tracing of the wireless radar in forming suspended point clouds.

[0030] Furthermore, S04 specifically includes the following steps: Obtain the seed attitude point cloud mesh of the wireless radar when reconstructing the smooth topology field from the source generated at the previous scan timestamp; Obtain the degrees of freedom of motion of the wireless radar itself or the detection carrier to which the wireless radar is attached, and construct a rigid skeleton of the wireless radar attitude based on the degrees of freedom of motion. Based on the attitude inversion projection points, the transformation implicit function of the rigid skeleton is defined. The influence of the attitude skeleton of different suspended point clouds in the suspended point cloud region is propagated and combined from the seed attitude point cloud mesh using the transformation implicit function to construct the dynamic transformation implicit space of the rigid skeleton. Based on the suspended vector, a transformation animation operation based on the degree of freedom of motion is applied to the rigid skeleton to obtain the motion transformation matrix of the rigid skeleton. The inverse mapping algorithm is introduced to inject the motion transformation matrix into the dynamic transformation implicit space to obtain the pre-transformation field value of each suspended point cloud. Based on the pre-transformation field value, each suspended point cloud is projected and propagated from the seed attitude point cloud mesh along the normal trajectory direction of the attitude inversion projection point to its corresponding suspended isosurface, thus obtaining the dynamic propagation matrix of attitude rigid transformation. Based on the dynamic propagation matrix, the point cloud positions of the seed attitude point cloud mesh are iteratively maintained and reconstructed through motion rigid transformation, and finally the current reconstructed attitude information of the wireless radar is output.

[0031] It should be noted that a rigid skeleton for wireless radar attitude control is constructed using motion degrees of freedom, and an implicit transformation function is defined for the rigid skeleton according to the attitude inversion projection points. This implicit transformation function represents the influence range of vertices of different suspended point clouds on the rigid skeleton, determining the attitude reconstruction driving contribution and responsibility share of each suspended point cloud following the motion characteristics of the wireless radar. Subsequently, transformation animation operations, including rotation and translation, are applied to the rigid skeleton to change the position and direction of different implicit fields in the rigid skeleton based on the suspended vectors, thereby customizing the target attitude and injecting the correct fitting driving force into the attitude reconstruction. Furthermore, the output motion transformation matrix is ​​applied to the dynamic transformation implicit space to query the spatial field value of each suspended point cloud before transformation, i.e., the pre-transformation field value. Since traditional forward transformation would destroy continuity, this method does not directly move vertices, ensuring that the field function remains smooth and does not produce attitude chain breaks. It avoids reconstruction artifacts caused by direct vertex linear interpolation, significantly improving the accuracy and reliability of attitude reconstruction estimation compared to traditional methods. This method can utilize the transformation information of suspended point clouds derived from deep learning to reconstruct attitude, thereby achieving a dynamic closed-loop effect of real-time attitude self-observation, updating, and maintenance for wireless radar.

[0032] A second aspect of the present invention provides a wireless radar attitude reconstruction device based on deep learning, such as... Figure 3 As shown, the device includes: a memory 301, a processor 302, and a communication interface 303. The memory 301 includes a deep learning-based wireless radar attitude reconstruction method program. The communication interface 303 is used for data connection communication between the memory 301 and the processor 302. When the wireless radar attitude reconstruction method program is executed by the processor 302, it implements any of the steps of the wireless radar attitude reconstruction method described above.

[0033] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A wireless radar attitude reconstruction method based on deep learning, characterized in that, Includes the following steps: S01: Acquire multiple frames of radar motion point cloud and wireless echo signal from the previous scan timestamp to the current scan timestamp of the wireless radar. Guide the local coherent structure propagation of the radar motion point cloud by the echo intensity of the attitude motion information contained in the wireless echo signal, complete the sparse point cloud motion, and obtain the dynamic dense point cloud set of the wireless radar motion. S02: Construct the topological structure of the smooth dynamic dense point cloud set. Reconstruct the smooth topological field of the target. Calculate the structural loss between the target smooth topological field and the source smooth topological field through tensor decomposition. Minimize the structural loss and align the dynamic point cloud structure of the motion. Under the premise of alignment, analyze the point cloud distribution with attitude change in the point cloud structure to determine the suspended point cloud region. S03: Apply the offset vector features of the moving point cloud in the source reconstruction smooth topology field using BEV bird's-eye view plane enhancement coding source and perform deep learning detection of the offset point cloud anchor box to obtain the view bounding box of the coarse motion attitude of the wireless radar. Interact the suspended point cloud area onto the surface projection backtracking of the source reconstruction smooth topology field with cross attention. Use the view bounding box as a coarse attitude constraint to drive the relative displacement of the surface projection to obtain the attitude inversion projection points and their suspended vectors of the suspended point cloud area. S04: Obtain the seed attitude point cloud mesh of the previous scan timestamp of the wireless radar, construct a rigid skeleton and its dynamic transformation implicit space based on the motion degrees of freedom of the wireless radar and the attitude inversion projection points, dynamically transform the rigid skeleton based on the suspended vector and inject it into the dynamic transformation implicit space, reconstruct the attitude rigid transformation of the suspended point cloud starting from the seed attitude point cloud mesh, and output the current reconstructed attitude information of the wireless radar.

2. The deep learning-based wireless radar attitude reconstruction method according to claim 1, characterized in that, S01 specifically includes the following steps: The system acquires multiple frames of radar motion point cloud continuously collected from the previous scan timestamp to the current scan timestamp during the time segment of the target scanning scene detected by the wireless radar, as well as the wireless echo signal corresponding to each frame of radar motion point cloud. At the same time, it acquires the predetermined sensing clock when the wireless radar outputs the motion echo signal. Based on the predetermined sensing clock, a motion sensing coordinate system of the wireless radar at the current scanning timestamp is constructed. A short-time Fourier transform algorithm is introduced to perform spatial frequency domain feature analysis on the wireless echo signal of each frame to obtain the signal transmit / receive vector, echo propagation time, frequency offset and echo array phase difference. Based on echo propagation time, frequency offset and echo array phase difference, the spatial transformation between radar moving point clouds in each frame is estimated. During the estimation process, the echo energy distribution contributing to the physical spatial location of each wireless echo signal is weighted according to the spatial transformation attribute to obtain the echo intensity guidance weight value. A weighted radial basis function field for spatial point cloud motion is constructed using PROE model software and multiple quadratic radial basis kernels. Based on echo intensity-guided weights, each wireless echo signal is introduced into the weighted radial basis function field to obtain a continuous radial basis coefficient vector for motion continuity. The coherent radial basis coefficient vector is used to interpolate and propagate the solution to the local homogeneous neighborhood of each radar motion point cloud in the structural direction, generating the sparse completion attributes and the probability distribution of the completion points for each frame of radar motion point cloud; wherein, the sparse completion attributes include the motion coordinates, spatial distance and structural representation level of the completion point cloud. Based on the sparse completion attribute and the probability distribution of the completion points, the radar motion point cloud of each frame is transformed and fixed on the motion sensing coordinate system for temporal fusion and registration, filling in the sparse and empty point cloud motion narrative, and obtaining the dynamic dense point cloud set of the wireless radar motion.

3. The deep learning-based wireless radar attitude reconstruction method according to claim 1, characterized in that, S02 specifically includes the following steps: By estimating the density characteristics of a dynamic dense point cloud, the elastic energy function of the radius neighborhood topology is defined based on the spatiotemporal Euclidean distance and dynamic integral iteration is performed based on it to construct a topological density attraction domain. The backbone topology is established by the topological connectivity partitions of the topological density attraction domain and the surrogate saddle points. The dynamic dense point cloud is smoothly distributed to the backbone topology to form a smooth topological field of point cloud motion of the dynamic dense point cloud, which is defined as the target reconstructed smooth topological field. Obtain the smoothed topological field of the point cloud motion at the previous scan timestamp of the wireless radar, and define it as the source reconstructed smoothed topological field; Tensor decomposition algorithm is introduced to decompose and calculate the source reconstructed smooth topological field and the target reconstructed smooth topological field respectively, to obtain the multidimensional structure loss tensor of the source reconstructed smooth topological field, which is labeled as the first structure loss tensor; and to obtain the multidimensional structure loss tensor of the target reconstructed smooth topological field, which is labeled as the second structure loss tensor. Calculate the structural loss gradient between the first structural loss tensor and the second structural loss tensor. Based on the structural loss gradient, preset the effective cost matrix of the motion structure of the target reconstructed smooth topology field relative to the source reconstructed smooth topology field. Align and match the point cloud structure of the source reconstructed smooth topology field to the target reconstructed smooth topology field. During the alignment matching process, the optimal matching transportation relationship of the point cloud structure is updated based on the effective cost matrix, and the structural hash degree of the target reconstructed smooth topology field relative to the source reconstructed smooth topology field is determined. The weighted covariance matrix between point cloud structures is calculated based on the criterion of minimizing structural scattering. A local point cloud distribution reference frame is constructed based on the weighted covariance matrix. Under the local point cloud distribution reference frame, the point cloud space is collaboratively divided into N sub-point cloud distribution domains. The angle between the normal vectors of different point clouds in the source reconstructed smooth topological field and the target reconstructed smooth topological field within each sub-point cloud distribution domain is calculated by histogram quantization and described and extracted, generating a suspended isosurface with attitude-invariant description in each sub-point cloud distribution domain. If the area of ​​the suspended isosurface is greater than the preset area value, then the sub-point cloud distribution domain is marked as a suspended point cloud region with attitude changes.

4. The deep learning-based wireless radar attitude reconstruction method according to claim 3, characterized in that, The process involves estimating the density characteristics of a dynamically dense point cloud, defining an elastic energy function for the radius neighborhood topology based on spatiotemporal Euclidean distance, and performing dynamic integral iterations to construct a topological density attraction domain. A backbone topological graph is then established using the topological connectivity partitions of the topological density attraction domain and surrogate saddle points. The dynamically dense point cloud is then smoothly distributed onto the backbone topological graph, forming a smooth topological field for the motion of the dynamically dense point cloud. Specifically, this includes the following steps: By constructing a spatial scalar potential field using PROE model software, and introducing a kernel density algorithm to estimate the continuous density and potential energy troughs of the input spatial scalar potential field of the dynamic dense point cloud, a density gradient vector field is obtained. Construct a radius neighborhood topology for a dynamic dense point cloud, obtain the time-varying spatiotemporal Euclidean distance between every two adjacent dynamic dense point clouds, and define the elastic energy function of the radius neighborhood topology based on the spatiotemporal Euclidean distance. The density gradient vector field is used to perform dynamic integral iteration on each dynamic dense point cloud in the dynamic dense point cloud set based on the elastic energy function, and the current local density value is output by dynamic maintenance. If the local density value has already converged to the center of the maximum local density value, then immediately stop the iterative operation of the dynamic integral, generate the structural convergence path of each dynamic dense point cloud corresponding to the topological center and the modal nodes along the path, and couple the structural convergence path of each dynamic dense point cloud to the modal nodes to construct the topological density attraction domain. The topological connectivity partitions and topological surrogate saddle points that change with time by topological density attraction domains are obtained. The main topological graph is jointly constructed with the topological surrogate saddle points as endpoints and the topological connectivity partitions as connecting edges. A continuous motion topology description code is compiled based on spatiotemporal Euclidean distance, and each dynamic dense point cloud is assigned and mapped onto the backbone topology map. The motion topology description code is then attached to the point cloud and smoothly distributed, ultimately forming a smooth motion topology field of the point cloud set.

5. The deep learning-based wireless radar attitude reconstruction method according to claim 1, characterized in that, S03 specifically includes the following steps: Introducing the BEV bird's-eye view plane, the source reconstructed smooth topological field is divided into k topological anchor pillars on the BEV bird's-eye view plane. The offset vectors of different moving point clouds within each topological anchor pillar are feature-enhanced encoded using a point cloud network, and the BEV pillar features of several candidate offset point clouds are output. Each BEV support feature is back-mapped to the predetermined position of the corresponding candidate offset point cloud on the BEV bird's-eye view plane to obtain the BEV pseudo offset support. The position branch and orientation branch of the BEV pseudo offset support are detected by CNN convolutional neural network and anchor-based anchor box detection method, and finally the view bounding box of the coarse motion attitude of the wireless radar is obtained. Each suspended point cloud in each of the suspended point cloud regions is interactively projected onto the motion structure surface represented by the source reconstructed smooth topological field. During the interactive projection process, the cross-attention information of the surface projection is simultaneously extracted from the source reconstructed smooth topological field to obtain the back-back projection representation of each suspended point cloud and its cross-attention weight. The back-projection representation includes the back-projection query vector, source key vector, and value vector of the suspended point cloud; The local surface features of each suspended point cloud region are aligned and aggregated based on the back-back projection representation to obtain the implicit attention surface. The relative displacement vector of the suspended point cloud is predicted using the implicit attention surface based on the cross attention weight, and the surface projection estimation result is generated. By introducing the source reconstruction smooth topological field as a geometric framework constraint using the view bounding box, the surface projection estimation results are projected along the coarse attitude gradient direction of the source reconstruction smooth topological field, thus obtaining the attitude inversion projection points and their dangling vectors of the dangling point cloud region.

6. The deep learning-based wireless radar attitude reconstruction method according to claim 1, characterized in that, S04 specifically includes the following steps: Obtain the seed attitude point cloud mesh of the wireless radar when reconstructing the smooth topology field from the source generated at the previous scan timestamp; Obtain the degrees of freedom of motion of the wireless radar itself or the detection carrier to which the wireless radar is attached, and construct a rigid skeleton of the wireless radar attitude based on the degrees of freedom of motion. Based on the attitude inversion projection points, the transformation implicit function of the rigid skeleton is defined. The influence of the attitude skeleton of different suspended point clouds in the suspended point cloud region is propagated and combined from the seed attitude point cloud mesh using the transformation implicit function to construct the dynamic transformation implicit space of the rigid skeleton. Based on the suspended vector, a transformation animation operation based on the degree of freedom of motion is applied to the rigid skeleton to obtain the motion transformation matrix of the rigid skeleton. The inverse mapping algorithm is introduced to inject the motion transformation matrix into the dynamic transformation implicit space to obtain the pre-transformation field value of each suspended point cloud. Based on the pre-transformation field value, each suspended point cloud is projected and propagated from the seed attitude point cloud mesh along the normal trajectory direction of the attitude inversion projection point to its corresponding suspended isosurface, thus obtaining the dynamic propagation matrix of attitude rigid transformation. Based on the dynamic propagation matrix, the point cloud positions of the seed attitude point cloud mesh are iteratively maintained and reconstructed through motion rigid transformation, and finally the current reconstructed attitude information of the wireless radar is output.

7. A wireless radar attitude reconstruction device based on deep learning, characterized in that, The device includes a memory, a processor, and a communication interface. The memory includes a deep learning-based wireless radar attitude reconstruction method program. The communication interface is used for data connection communication between the memory and the processor. When the wireless radar attitude reconstruction method program is executed by the processor, it implements the steps of the wireless radar attitude reconstruction method as described in any one of claims 1-6.

Images