Millimeter wave synthetic aperture radar imaging method and system based on quadruped robot

Abstract

The invention relates to the technical field of mobile robot platform radar imaging, in particular to a millimeter wave synthetic aperture radar imaging method and system based on a quadruped robot. The method comprises the steps that a traditional horizontal scanning track is adjusted into a vertical scanning path, and the quadruped robot is controlled to execute cyclic vertical and horizontal compound motion so as to construct a virtual antenna array; constructing a pose map by using sensor data of the robot, estimating a motion track through a pose map optimization algorithm, and compensating a phase error; and sequentially applying a frequency domain filtering fringe removing method and a generative adversarial network image restoration method taking an optical image as a reference to the initial SAR image to reconstruct a target contour and restore a missing region. According to the method, the influence of non-rigid motion of the quadruped robot on imaging is overcome, robust high-resolution imaging under shielding and non-line-of-sight scenes is realized, and the complex environment sensing capability of a mobile robot platform is expanded.

Description

Technical Field

[0001] This invention relates to the field of radar imaging technology for mobile robot platforms, and more specifically to a millimeter-wave synthetic aperture radar imaging method and system based on quadruped robots. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging radar that utilizes the motion of radar platforms (satellites, aircraft, etc.) and complex signal processing techniques to synthesize small-sized real antennas into large-sized equivalent "virtual antennas," thereby obtaining high-resolution two-dimensional or three-dimensional images. It has become a powerful and versatile sensing technology. By synthesizing a large antenna aperture, SAR radar overcomes the physical limitations of traditional radar hardware, achieving high-resolution sensing, and its applications range from satellite remote sensing to various civilian radar imaging scenarios.

[0003] Millimeter waves have a certain penetrating ability, enabling them to sense objects while penetrating obstacles, and therefore have received increasing attention in various sensing fields.

[0004] Millimeter-wave radar (mmWave Radar) is an active electromagnetic sensing technology that operates in the millimeter-wave frequency band (30GHz~300GHz, corresponding to wavelengths of 1mm~10mm). By emitting high-frequency electromagnetic waves and analyzing the echo signals, it can accurately measure the three-dimensional information of a target, including its distance, velocity, and angle. With the continuous advancement of its technology, millimeter-wave SAR radar suitable for short-range, high-resolution sensing has been developed.

[0005] The millimeter-wave SAR radar is a fusion of millimeter-wave technology and synthetic aperture radar technology. It combines the advantages of millimeter waves, such as short wavelength, high bandwidth, and miniaturization, with the virtual large aperture and high azimuth resolution capabilities of SAR. It is one of the cutting-edge directions in the field of radar technology. The millimeter-wave SAR radar is essentially a synthetic aperture radar system operating in the 30GHz~300GHz (wavelength 1mm~10mm) millimeter-wave frequency band.

[0006] Currently, millimeter-wave SAR radar typically relies on the movement of mechanical rails to simulate the antenna displacement required for aperture synthesis. While these systems provide accurate and stable apertures, their mechanical components are static, making them unsuitable for mobile scenarios. Other research has explored building radar platforms using different carriers, such as automobiles, handheld devices, and drones. However, these methods have limitations in terms of operational complexity and application scope. These limitations have led to a need for more flexible and adaptable SAR radar solutions for various scenarios.

[0007] Quadruped robots are gaining popularity in military, industrial, and civilian fields due to their mobility and environmental adaptability. Existing work has integrated millimeter-wave radar with quadruped robots, but this typically involves rapidly rotating the millimeter-wave radar like a lidar system. This method can only generate general environmental overview images and cannot provide the object-level imaging resolution required to identify fine structural details. Furthermore, it relies on custom-made rotary motors, increasing hardware complexity and limiting potential application scenarios.

[0008] Based on the aforementioned existing technologies, realizing millimeter-wave synthetic aperture radar (SAR) using quadruped robot motion requires addressing the following practical challenges.

[0009] 1. Design challenges of synthetic aperture for robot motion: Due to the physical structure and motion patterns of quadruped robots, their synthetic aperture methods differ significantly from traditional linear track-based SAR radar. Therefore, it is necessary to design a specific trajectory and scanning method that fits the motion characteristics of quadruped robots to achieve effective virtual aperture synthesis.

[0010] 2. Compensation for Robot Motion Errors: Utilizing the non-rigid, multi-joint motion of a quadruped robot for aperture synthesis introduces various motion errors, such as vibration, trajectory deviation, and attitude angle changes. These errors cause phase errors in the radar echo signal. Without effective compensation, the accuracy of the synthesized aperture will be severely affected, potentially leading to imaging failure or a significant performance degradation.

[0011] 3. Achieving High-Quality Imaging: Although large apertures theoretically provide finer-grained sensing, the actual imaging process based on millimeter-wave radar is limited by signal capacity and inherent noise. Initial imaging results often suffer from problems such as striped artifacts, distortion, and missing information. Therefore, subsequent image processing methods are needed to suppress artifacts and repair images to improve the practicality, clarity, and usability of the final imaging results in complex environments. Summary of the Invention

[0012] In view of this, the present invention provides a millimeter-wave synthetic aperture radar imaging method and system based on a quadruped robot, which aims to solve the three core problems faced by existing technologies in achieving high-resolution millimeter-wave SAR imaging using a quadruped robot platform: synthetic aperture construction, motion error compensation, and image quality improvement, thereby enhancing the quadruped robot's fine perception and robust imaging capabilities in complex scenarios such as occlusion and non-line-of-sight.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a millimeter-wave synthetic aperture radar imaging method based on a quadruped robot, comprising: The horizontal scanning trajectory is adjusted to a vertical scanning path, and the quadruped robot is controlled to perform cyclical vertical and horizontal compound motion, so that the onboard millimeter-wave radar moves along a specific trajectory to construct a virtual antenna array. A pose map is constructed using the robot's own sensor data. The motion trajectory is estimated through a pose map optimization algorithm. The phase error is calculated based on the deviation between the estimated motion trajectory and the ideal motion trajectory, and phase compensation is performed on the radar echo signal. An initial SAR image is obtained by imaging the phase-compensated radar echo signal. The initial SAR image is then sequentially processed by a frequency domain filtering destriating method and a generative adversarial network image restoration method trained with optical images as real labels to reconstruct the target contour and repair missing regions, resulting in an enhanced SAR image.

[0014] In one specific implementation scheme, controlling the quadruped robot to perform cyclical vertical and horizontal compound movements specifically includes: Control the quadruped robot to stand upright with its limbs at the highest point, then bend its limbs to descend to the lowest point, and then return to a natural standing position. After moving horizontally a preset distance, repeat the above vertical movement until the set cycle endpoint is reached. The vertical motion sampling interval is set to 2mm, the horizontal movement interval is set to 8mm, the total horizontal movement length is 240mm, the total vertical movement length is 130mm, the duration of a single vertical movement is 1.3s, and the sampling time interval is set to 0.02s.

[0015] In one specific implementation scheme, the step of constructing a pose graph using the robot's own sensor data and estimating the motion trajectory using a pose graph optimization algorithm specifically includes: The robot pose at each sampling point. As nodes in the pose graph, For robots in Position coordinates at that moment; Displacement between adjacent nodes As edges in the pose graph; Speed ​​information measured using a speed sensor Calculate the predicted pose ; Define error function By minimizing the objective function Solve for the optimal pose set ,in For information matrix, The standard deviations in the x, y, and z directions are all set to 0.005.

[0016] In one specific implementation scheme, the step of calculating the phase error based on the deviation between the estimated motion trajectory and the ideal motion trajectory, and performing phase compensation on the radar echo signal, specifically includes: Based on the robot's pose at time t and the fixed offset between the radar and the robot body Pitch angle Roll angle Calculate the actual coordinates of the radar for: The pitch and roll angles are obtained by the robot's own accelerometers. Calculate the distance error between the actual radar coordinates and the ideal position at time t. ,in Let be the ideal position vector of the radar at time t. = This is the actual position vector; Calculate phase change based on distance error ,in The wavelength of the millimeter-wave signal; Using formula radar echo signal Phase correction is performed to obtain the compensated signal. .

[0017] In one specific implementation, the step of imaging the phase-compensated radar echo signal to obtain an initial SAR image specifically includes: processing the compensated radar echo signal using the Range Migration Algorithm (RMA) to obtain the initial SAR image.

[0018] In one specific implementation scheme, the frequency domain filtering destriping method specifically includes: For the initial SAR image The frequency domain spectrum is obtained by applying the Fast Fourier Transform (FFT). ,in Image size, Frequency coordinates; Constructing a frequency domain mask Set the center horizontal frequency component to 0, and keep the other frequency components at 1; Calculate the filter spectrum ; Performing an inverse fast Fourier transform (IFFT) on the filtered spectrum yields the destriped image. .

[0019] In one specific implementation, the generative adversarial network trained with optical images as real labels has a generator... An encoder-decoder architecture is adopted. The encoder consists of 6 downsampling layers with 4×4 convolutional kernels and a stride of 2. The decoder consists of the same number of upsampling layers. Feature concatenation is performed between corresponding layers of the encoder and decoder through skip connections.

[0020] In one specific implementation scheme, the generative adversarial network is trained in the following manner: The input noise vector z is composed of the initial SAR image and the corresponding optical image; The objective function to be optimized is: in The optical image corresponding to the object; The loss function consists of adversarial loss and L1 loss, with the weight λ of the L1 loss set to 100. The Adam optimizer is used, with a learning rate of 0.0002 and a first-order moment decay coefficient. Second-order moment attenuation coefficient The batch size was set to 64, and a total of 200 epochs were trained.

[0021] Secondly, the present invention provides a millimeter-wave synthetic aperture radar imaging system based on a quadruped robot, the system being used to implement the aforementioned millimeter-wave synthetic aperture radar imaging method based on a quadruped robot, comprising: The quadruped robot platform, equipped with millimeter-wave radar, is configured to perform cyclical vertical and horizontal compound motions, causing the millimeter-wave radar to move along a specific trajectory to construct a virtual antenna array. The motion compensation module is configured to construct a pose map using the robot's own sensor data, estimate the motion trajectory through a pose map optimization algorithm, calculate the phase error based on the deviation between the estimated motion trajectory and the ideal motion trajectory, and perform phase compensation on the radar echo signal. The image processing module is configured to perform imaging processing on the phase-compensated radar echo signal to obtain an initial SAR image, and then apply a frequency domain filtering destriating method and a generative adversarial network image restoration method trained with optical images as real labels to the initial SAR image to reconstruct the target contour and repair missing regions, thereby obtaining an enhanced SAR image.

[0022] In one specific implementation, the system further includes sensors for acquiring robot pose data, including a position sensor, a velocity sensor, and an accelerometer; the motion compensation module and the image processing module are integrated into the processing unit of the quadruped robot platform.

[0023] Compared with existing technologies, the millimeter-wave synthetic aperture radar imaging method and system based on a quadruped robot described in this invention is used to achieve high-resolution perception in complex scenes using a quadruped robot platform. By designing a scanning trajectory adapted to the robot's motion characteristics, utilizing a motion error compensation method based on pose graph optimization, and combining image artifact removal and generative adversarial networks for image restoration, the entire process from radar data acquisition and signal correction to image enhancement is realized, effectively improving the imaging accuracy, robustness, and practicality of millimeter-wave synthetic aperture radar in unstructured environments, and has the following beneficial effects: 1. An effective synthetic aperture (SAR) construction adapted to the motion characteristics of a quadruped robot was achieved. By adjusting the traditional horizontal scanning trajectory to a vertical (gravity direction) scanning path and controlling the robot to perform cyclical vertical and horizontal composite movements, the millimeter-wave radar can move along a specific trajectory, thereby constructing a virtual antenna array suitable for SAR imaging. This method overcomes the limitations of the quadruped robot's physical structure and motion mode on the SAR, providing a feasible aperture synthesis scheme for achieving high-quality SAR imaging using a mobile robot platform.

[0024] 2. Significantly reduces the negative impact of robot motion errors on imaging quality. By utilizing the robot's own sensor data, the motion process is modeled as a pose graph, and a phase correction method based on pose graph optimization is employed to accurately estimate and compensate for phase errors caused by robot vibration, trajectory deviation, and attitude angle changes. This technique maps radar measurements from the interfered actual motion path to the ideal position, effectively correcting the echo signal and providing high-quality input data for subsequent high-precision imaging processing.

[0025] 3. Improved clarity and integrity of the final SAR image. By applying a frequency-domain filtering-based image destriating method to suppress periodic artifacts, and further employing a generative adversarial network-based image restoration method to reconstruct target contours and repair missing regions, interference and distortion during the imaging process can be effectively reduced. This combination of image processing and deep learning techniques enables the system to output clearer, more accurate, and usable SAR images, enhancing its practical value in complex sensing scenarios. Attached Figure Description

[0026] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0027] Figure 1This is an overall flowchart of the millimeter-wave synthetic aperture radar imaging method based on a quadruped robot described in this invention.

[0028] Figure 2 This is an architectural diagram of the millimeter-wave synthetic aperture radar imaging system based on a quadruped robot as described in this invention.

[0029] Figure 3 Design of the motion trajectory for a quadruped robot.

[0030] Figure 4 This is the implementation method of RobSAR.

[0031] Figure 5 The following objects were SAR images taken in an accessible scene: (a) a round stainless steel cup lid, (b) a rectangular portable hard drive, (c) a square plate, (d) an oval box, (e) a slender metal hammer, and (f) an electric drill.

[0032] Figure 6 SAR imaging results for different objects obscured by different obstacles: (a) is a round stainless steel cup lid, (b) is a rectangular portable hard drive, (c) is a square plate, (d) is an oval box, (e) is a slender metal hammer, and (f) is an electric drill. Detailed Implementation

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] This embodiment provides a millimeter-wave synthetic aperture radar imaging method based on a quadruped robot, including: The horizontal scanning trajectory is adjusted to a vertical scanning path, and the quadruped robot is controlled to perform cyclical vertical and horizontal compound motion, so that the onboard millimeter-wave radar moves along a specific trajectory to construct a virtual antenna array. A pose map is constructed using the robot's own sensor data. The motion trajectory is estimated through a pose map optimization algorithm. The phase error is calculated based on the deviation between the estimated motion trajectory and the ideal motion trajectory, and phase compensation is performed on the radar echo signal. An initial SAR image is obtained by imaging the phase-compensated radar echo signal. The initial SAR image is then sequentially processed by a frequency domain filtering destriating method and a generative adversarial network image restoration method trained with optical images as real labels to reconstruct the target contour and repair missing regions, resulting in an enhanced SAR image.

[0035] The method described in this invention is based on a millimeter-wave synthetic aperture radar imaging system using a quadruped robot. Figure 2 This novel SAR imaging system, named RobSAR, utilizes the motion of a quadruped robot to synthesize a large virtual aperture using millimeter-wave radar. This enables the quadruped robot to operate through obstacle imaging, opening up new opportunities for safety-critical and obstructed scenarios such as rescue and inspection. Figure 1 The overall process of the method described in this invention will be explained in detail below with reference to the accompanying drawings.

[0036] To achieve the above objectives, the specific implementation scheme of the present invention is described as follows: 1. Synthetic aperture of a quadruped robot To achieve SAR imaging using millimeter-wave radar and a quadruped robot, a synthetic virtual antenna array is required. By simulating an antenna array larger than the actual physical antenna, synthetic aperture radar can provide high-quality sensing for millimeter-wave radar. To construct the synthetic aperture virtual antenna, a sensing model of millimeter-wave synthetic aperture radar was established. The radar's position is available as it moves in three-dimensional space. This indicates that n is the index of the sampling point of the received signal. The sensing target is located at... For a single antenna, the distance between the target and the antenna can be calculated using Euclidean distance. Therefore, the echo signal received by the antenna can be expressed as: in, It is the amplitude coefficient. It is the wavelength of the millimeter-wave signal.

[0037] The echo data is the sum of the echo signals received by the antenna at each sampling point, i.e., the signals from all virtual antennas. Therefore, the echo data of the virtual antenna array can be represented as: Because the displacement of the radar platform is limited by the robot's shape, the size of the virtual antenna array is restricted. Therefore, it is necessary to further design the radar platform motion to conform to the quadruped robot's shape to achieve aperture synthesis. To synthesize an aperture, the radar needs to slide along a specific track to scan an area. In traditional radar scanning, millimeter-wave radar first moves horizontally, then vertically, and repeats the horizontal movement. With the help of a two-dimensional sliding track, the scanning process constructs an aperture... The two-dimensional scanning plane. Radar sliding. The spacing between two adjacent rows is . Collect each line There are 1 sampling point, with an interval of 1. .

[0038] Although rail-based radar scanning is simple, it is difficult to implement a similar scanning process using a quadruped robot due to motion limitations. To accommodate the mobility of quadruped robots, this invention adjusts the horizontal SAR scanning trajectory to a vertical scanning path, such as... Figure 3 As shown, a quadruped robot equipped with millimeter-wave radar stands upright, positioning itself at its highest point. It then bends its limbs to reach its lowest point. Afterward, it returns to its natural standing position and moves horizontally a short distance. This cycle of vertical and horizontal movement continues until a predetermined endpoint is reached. By selecting sampling points, a vertical scanning configuration achieves similar performance, supporting SAR imaging. This invention sets the vertical sampling interval to 2 mm and the horizontal movement interval to 8 mm. The total lateral movement length is 240 mm, requiring 30 repetitions of this action. The total vertical movement length is 130 mm, and the duration of a single vertical movement, from the highest to the lowest point, is 1.3 s, meaning the sampling time interval is set to 1.3 / 65 s = 0.02 s.

[0039] 2. Robot motion compensation Precise motion control is crucial for the synthetic aperture radar (SAR) of quadruped robots. Unlike traditional rail-based SAR, quadruped robot-based SAR is affected by motion vibrations and trajectory deviations caused by the mechanical structure. However, maintaining an ideal motion path in real-world environments using quadruped robots as mobile radar platforms presents significant challenges. This is due to the limitations of controlling the quadruped robot's mechanical structure in complex terrain.

[0040] In the vertical movement of quadruped robots, the horizontal stepping accuracy is low, making them prone to deviations. This leads to errors along the horizontal axis. Quadruped robots move vertically by extending their limbs, a non-uniform motion. This results in unequal distances between sampling points along the vertical line. Complex real-world environments also introduce other interferences. When a quadruped robot moves laterally and traverses complex terrain, the different heights of its four legs cause body yaw. While the robot can adjust its flexion when height changes are small, it cannot maintain horizontal alignment under large height changes. For example, when the robot's left foot is one step higher than its right, its body tilts to the right, causing a roll angle deviation. When the robot's front foot is one step higher than its rear foot, its body tilts upward, causing a pitch angle deviation. These errors result in an angle between the actual movement path and the vertical line. Positional deviations can lead to significant phase errors, potentially causing SAR imaging failures or degraded sensing performance.

[0041] To reduce the impact of robot motion errors, a precise motion compensation method is needed. This invention utilizes the robot's own position and velocity sensors to obtain its motion state. The position sensor has an error of 0.001m when stationary and 0.005m when moving. The velocity sensor has an error of 0.001m / s. Assuming the robot's body position is... The fixed offset between the radar and the robot body is This fixed offset is the offset between the millimeter-wave radar position and the robot's position when the robot is standing normally. This value is a fixed precision of 19 cm. The pitch angle is... The roll angle is The pitch and roll angles are both obtained from the robot's own accelerometers, with a measurement accuracy of 0.1°. Therefore, the radar's global coordinates can be expressed as... Due to the motion interference of the quadruped robot, the actual coordinates of the radar differ from the ideal position, as shown below. .

[0042] The purpose of motion compensation is to map radar measurements from their actual motion path to their corresponding ideal position to eliminate the effects of motion interference. For quadruped robots, error correction needs to be performed after the signal is received. To this end, this invention proposes a phase correction method based on pose graph optimization (PGO) to compensate for motion errors in quadruped robots.

[0043] The goal of motion compensation is to correct the phase of the acquired signal. Phase correction converts the signal into one acquired at an ideal location, thus eliminating the influence of motion errors. Assume... This represents the target's coordinates relative to the actual radar coordinates at time t. The distance between them This represents the target's coordinates relative to the ideal radar coordinates at time t. The distance between them. Since the difference between the two coordinates is much smaller than the distance between the radar and the target, the distance error over time t can be approximated as... The phase of the radar echo signal is Distance change Cause phase change .

[0044] This invention treats each sampling point as a pose and the entire motion process as a graph. By optimizing the pose graph of the quadruped robot, the motion trajectory of the quadruped robot is solved, accurately compensating for the robot's motion errors. Assume that at time t, the robot's position is... The speed is .in The robot's velocity values ​​along the x, y, and z axes are all directly measured by sensors. The robot's motion process is modeled graphically. In the figure, this invention represents time... The pose is set as a node ,and Set the edges of the graph to the displacements between adjacent poses. Since the overall direction of velocity is precise, velocity information is used to predict position and estimate trajectory. The robot performs a 130mm vertical scan over a fixed duration of 1.3s. With a vertical sampling interval of 2mm, meaning 65 samples are completed within 1.3s, the sampling frequency is 65 / 1.3 = 50Hz. Synchronization is achieved by triggering radar acquisition the instant the robot begins its downward movement, selecting one radar frame every 20ms. This scheme ensures that each radar frame corresponds to a known robot pose along the vertical trajectory. Using the velocity information, the predicted pose can be obtained as follows: Each sampling point of the quadruped robot can be obtained based on the error function. The error function is as follows: , which represents the positional error between the measured pose and the predicted pose.

[0045] For each vertical movement of the quadruped robot, the motion path is obtained by optimizing the pose map. Since the robot can achieve accurate position measurement while stationary, the highest and lowest points during vertical movement can be accurately measured. Therefore, maintaining... and The goal of PGO remains unchanged. It is the set of all poses. The information matrix is ​​a key parameter describing the uncertainty of constraints. This represents the standard deviation in the x, y, and z directions. Since the three directions have the same impact on the robot's position, and the robot's positional error during movement is 5mm, the standard deviation is set to... Next, this invention uses a generalized graph optimization (g2o) framework to obtain the minimum value.

[0046] After obtaining the robot's motion path, the robot's position at time t can be converted into radar coordinates. Therefore, the distance can be calculated. With phase Then you can use the formula. Phase correction is then performed. Finally, using the compensated radar echo signal, the range migration algorithm (RMA) is employed to achieve initial SAR imaging.

[0047] 3. SAR Image Denoising and Restoration Due to signal imperfections, the currently generated images exhibit stripes and distortion. To further improve imaging quality, this invention employs image denoising and image inpainting methods to support quadrupedal robot SAR.

[0048] To address the stripe artifacts in SAR images, this invention employs an image destriping method. In this method, the invention applies a Fast Fourier Transform (FFT) to the original SAR image, converting the spatial domain signal into a frequency domain spectrum in the following manner: in, The original SAR image, For frequency coordinates, Image size, It represents the frequency spectrum.

[0049] In the spectrum, periodic fringes in the spatial domain appear as symmetrical bright spots, typically perpendicular to the fringe direction. To suppress these fringe artifacts, this invention constructs a frequency domain mask. This is used to selectively attenuate the corresponding frequency components. In SAR images, stripe noise often appears in a vertical pattern. Therefore, in the mask, the center horizontal frequency component is set to 0, while other frequency components remain at 1. The filtered spectrum is calculated as follows: in These are the spatial coordinates of the output image. For frequency coordinates, Image size The filtered spectrum is calculated as follows: .

[0050] Next, based on the fundamental ideas of Generative Adversarial Networks (GANs), this invention designs a neural network model to enhance generated images. The optimization can be formalized as a minimax game with the following objective function: In this formula, the first term represents correctly identifying the real sample. The expected value, the second term represents the incorrectly generated sample. The expected value. Distribution and These represent the probabilities of sampling from real and generated datasets, respectively. In this invention, The optical image corresponds to the object, obtained by placing the object against a white background and taking a photograph. The optical image is used as the ground truth. Input noise vector. The generator combines a SAR image with an optical image of the target. Both the SAR and optical images of the same object are input into the generator. The generator uses the SAR image as a base and adjusts the original SAR image according to the shape of the object in the optical image to produce a high-quality SAR image. Enhanced SAR images are obtained using a traditional encoder-decoder architecture. The generator G has a symmetrical architecture. The encoder consists of six downsampling layers, each using a 4×4 convolutional kernel with a stride of 2 for spatial dimensionality reduction, doubling the number of channels layer by layer, and employing LeakyReLU as the activation function to preserve gradient information. The decoder gradually restores spatial resolution through the same number of upsampling layers, using the ReLU activation function. Feature concatenation is performed between corresponding layers of the encoder and decoder via skip connections to preserve low-level details. The network uses instance normalization layers for stable training, and the final output layer uses the sigmoid activation function to map pixel values ​​to the target area. Then, the discriminator... The shapes of optical and SAR images are obtained by detecting color. In near-field SAR images, the color depth gradually decreases from red to blue from the image center to the periphery; therefore, the shape of the SAR image can be obtained by setting the color value range. The optical image used is an object placed against a white background; its shape can also be obtained using the same method. During training, we use the Adam optimizer with a learning rate of 0.0002 and a first-order moment decay coefficient. Second-order moment attenuation coefficient The batch size was set to 64, and training lasted for 200 epochs. The loss function consisted of two parts: adversarial loss and L1 loss. The adversarial loss adopted the standard GAN loss form, and the L1 loss weight λ was set to 100 to enhance the structural similarity between the generated and real images. All network weights were randomly initialized from a normal distribution with a mean of 0 and a standard deviation of 0.02. Then, the SAR image was identified as either a real image or an image generated by a network. The generated image is output, along with the probability that the image is real. This problem is formulated as a zero-sum adversarial game until... It cannot distinguish between real and generated images. This indicates that... With sufficient training, reasonable images can be generated. By employing a GAN-based image restoration method, this invention reconstructs accurate target contours from real images. This method enhances the practicality and usability of the generated SAR images in real-world applications.

[0051] Once training is complete, in the practical application phase, low-quality SAR images can be directly input to obtain corresponding high-quality SAR images. In the application phase, only the generator is needed. First, the model is switched to evaluation mode, allowing the instance normalization layer to use global statistics from training and disabling gradient computation to improve inference efficiency. The input SAR image undergoes the same preprocessing as in the training phase, then performs a forward propagation through the generator to obtain the output tensor. Finally, the pixel values ​​are mapped to the [0,1] interval using the sigmoid activation function of the output layer and denormalized to the standard image format, thus obtaining the final high-quality SAR image.

[0052] By employing image destriating methods and GAN-based image restoration, this invention effectively reduces interference artifacts and improves the quality of SAR images.

[0053] The above method will be further illustrated by specific embodiments below.

[0054] This invention implements a RobSAR system using the Unitree Go2 quadruped robot. A Texas Instruments IWR1843BOOST millimeter-wave radar is connected to a DCA1000EVM data acquisition card, which is placed on top of the quadruped robot, as shown below. Figure 4 As shown. The placement method is as follows: First, place the millimeter-wave radar on two matching metal mounting brackets. The width of the mounting brackets on the millimeter-wave radar is the same as the width of the expansion dock on top of the quadruped robot. Therefore, use screws to fix it to the expansion dock. The distance between the top of the metal bracket and the top of the expansion dock is 5cm. After fixing, the millimeter-wave radar itself will not jitter when the robot moves. The millimeter-wave radar has three transmitting antennas and four receiving antennas. The millimeter-wave radar is configured to transmit a linear frequency modulated signal between 77 and 81 GHz, and a single pulse signal within 50 ms. The sampling rate is 7000 ksps. This invention controls the robot's movement based on the analysis and design in the synthetic aperture of the quadruped robot. The robot first rises from its original height to a maximum height of 30 mm, then descends to a minimum height of -100 mm and returns to its original height. The lateral and longitudinal sampling intervals are 8 mm and 2 mm, respectively.

[0055] During training, 20 sets of data were collected for each imaged object. The training data collection process was conducted in an unobstructed environment. The training data included sampled millimeter-wave data along a pre-selected SAR motion trajectory, as well as images of the target. During the process, 5 sets of data were collected for each object to evaluate imaging capabilities. The evaluation included both unobstructed and obstructed scenarios. In the obstructed scenario, four obstacles were used: a piece of paper, a cardboard box, a plastic bag, and a piece of clothing. These are common items in everyday environments that may become obstacles for the robot's perception in certain situations. Obstacles were placed between the quadruped robot and the target object. The robot's motion and data acquisition methods were the same as in the unobstructed scenario. Imaging results are as follows: Figure 5 and Figure 6 As shown.

[0056] The above experiments verified the effectiveness of the present invention. Extensive real-world experiments show that the RobSAR system can achieve robust and accurate SAR imaging in both line-of-sight and non-line-of-sight environments.

[0057] This embodiment provides a millimeter-wave synthetic aperture radar imaging system based on a quadruped robot. The system is used to implement the aforementioned millimeter-wave synthetic aperture radar imaging method based on a quadruped robot, including: The quadruped robot platform, equipped with millimeter-wave radar, is configured to perform cyclical vertical and horizontal compound motions, causing the millimeter-wave radar to move along a specific trajectory to construct a virtual antenna array. The motion compensation module is configured to construct a pose map using the robot's own sensor data, estimate the motion trajectory through a pose map optimization algorithm, calculate the phase error based on the deviation between the estimated motion trajectory and the ideal motion trajectory, and perform phase compensation on the radar echo signal. The image processing module is configured to perform imaging processing on the phase-compensated radar echo signal to obtain an initial SAR image, and then apply a frequency domain filtering destriating method and a generative adversarial network image restoration method trained with optical images as real labels to the initial SAR image to reconstruct the target contour and repair missing regions, thereby obtaining an enhanced SAR image.

[0058] The system also includes sensors for acquiring robot pose data, including position sensors, velocity sensors, and accelerometers; the motion compensation module and image processing module are integrated into the processing unit of the quadruped robot platform.

[0059] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A millimeter-wave synthetic aperture radar imaging method based on a quadruped robot, characterized in that, include: The horizontal scanning trajectory is adjusted to a vertical scanning path, and the quadruped robot is controlled to perform cyclical vertical and horizontal compound motion, so that the onboard millimeter-wave radar moves along a specific trajectory to construct a virtual antenna array. A pose map is constructed using the robot's own sensor data. The motion trajectory is estimated through a pose map optimization algorithm. The phase error is calculated based on the deviation between the estimated motion trajectory and the ideal motion trajectory, and phase compensation is performed on the radar echo signal. An initial SAR image is obtained by imaging the phase-compensated radar echo signal. The initial SAR image is then sequentially processed by a frequency domain filtering destriating method and a generative adversarial network image restoration method trained with optical images as real labels to reconstruct the target contour and repair missing regions, resulting in an enhanced SAR image.

2. The millimeter-wave synthetic aperture radar imaging method based on a quadruped robot according to claim 1, characterized in that, The control of the quadruped robot to perform cyclical vertical and horizontal compound movements specifically includes: Control the quadruped robot to stand upright with its limbs at the highest point, then bend its limbs to descend to the lowest point, and then return to a natural standing position. After moving horizontally a preset distance, repeat the above vertical movement until the set cycle endpoint is reached. The vertical motion sampling interval is set to 2mm, the horizontal movement interval is set to 8mm, the total horizontal movement length is 240mm, the total vertical movement length is 130mm, the duration of a single vertical movement is 1.3s, and the sampling time interval is set to 0.02s.

3. The millimeter-wave synthetic aperture radar imaging method based on a quadruped robot according to claim 1, characterized in that, The process of constructing a pose graph using the robot's own sensor data and estimating the motion trajectory using a pose graph optimization algorithm specifically includes: The robot pose at each sampling point. As nodes in the pose graph, For robots in Position coordinates at that moment; Displacement between adjacent nodes As edges in the pose graph; Speed ​​information measured using a speed sensor Calculate the predicted pose ； Define error function By minimizing the objective function Solve for the optimal pose set ,in For information matrix, The standard deviations in the x, y, and z directions are all set to 0.

005.

4. The millimeter-wave synthetic aperture radar imaging method based on a quadruped robot according to claim 3, characterized in that, The step of calculating the phase error based on the deviation between the estimated trajectory and the ideal trajectory, and performing phase compensation on the radar echo signal, specifically includes: Based on the robot's pose at time t and the fixed offset between the radar and the robot body Pitch angle Roll angle Calculate the actual coordinates of the radar for: The pitch and roll angles are obtained by the robot's own accelerometers. Calculate the distance error between the actual radar coordinates and the ideal position at time t. ,in Let be the ideal position vector of the radar at time t. = This is the actual position vector; Calculate phase change based on distance error ,in The wavelength of the millimeter-wave signal; Using formula radar echo signal Phase correction is performed to obtain the compensated signal. .

5. The millimeter-wave synthetic aperture radar imaging method based on a quadruped robot according to claim 1, characterized in that, The process of imaging the phase-compensated radar echo signal to obtain the initial SAR image specifically includes: using the Range Migration Algorithm (RMA) to process the compensated radar echo signal to obtain the initial SAR image.

6. The millimeter-wave synthetic aperture radar imaging method based on a quadruped robot according to claim 1, characterized in that, The frequency domain filtering destriping method specifically includes: For the initial SAR image The frequency domain spectrum is obtained by applying the Fast Fourier Transform (FFT). ,in Image size, Frequency coordinates; Constructing a frequency domain mask Set the center horizontal frequency component to 0, and keep the other frequency components at 1; Calculate the filter spectrum ; Performing an inverse fast Fourier transform (IFFT) on the filtered spectrum yields the destriped image. .

7. The millimeter-wave synthetic aperture radar imaging method based on a quadruped robot according to claim 1, characterized in that, The generative adversarial network trained using optical images as real labels has a generator. An encoder-decoder architecture is adopted. The encoder consists of 6 downsampling layers with 4×4 convolutional kernels and a stride of 2. The decoder consists of the same number of upsampling layers. Feature concatenation is performed between corresponding layers of the encoder and decoder through skip connections.

8. The millimeter-wave synthetic aperture radar imaging method based on a quadruped robot according to claim 7, characterized in that, The generative adversarial network is trained in the following manner: The input noise vector z is composed of the initial SAR image and the corresponding optical image; The objective function to be optimized is: ； in The optical image corresponding to the object; The loss function consists of adversarial loss and L1 loss, with the weight λ of the L1 loss set to 100. The Adam optimizer is used, with a learning rate of 0.0002 and a first-order moment decay coefficient. Second-order moment attenuation coefficient The batch size was set to 64, and a total of 200 epochs were trained.

9. A millimeter-wave synthetic aperture radar imaging system based on a quadruped robot, characterized in that, The system is used to implement the millimeter-wave synthetic aperture radar imaging method based on a quadruped robot as described in any one of claims 1-8, comprising: The quadruped robot platform, equipped with millimeter-wave radar, is configured to perform cyclical vertical and horizontal compound motions, causing the millimeter-wave radar to move along a specific trajectory to construct a virtual antenna array. The motion compensation module is configured to construct a pose map using the robot's own sensor data, estimate the motion trajectory through a pose map optimization algorithm, calculate the phase error based on the deviation between the estimated motion trajectory and the ideal motion trajectory, and perform phase compensation on the radar echo signal. The image processing module is configured to perform imaging processing on the phase-compensated radar echo signal to obtain an initial SAR image, and then apply a frequency domain filtering destriating method and a generative adversarial network image restoration method trained with optical images as real labels to the initial SAR image to reconstruct the target contour and repair missing regions, thereby obtaining an enhanced SAR image.

10. The millimeter-wave synthetic aperture radar imaging system based on a quadruped robot according to claim 9, characterized in that, It also includes sensors for collecting robot pose data, including position sensors, velocity sensors and accelerometers; the motion compensation module and image processing module are integrated into the processing unit of the quadruped robot platform.

Images