Millimeter wave radar and vision fusion-based swimming pool drowning prevention early warning system and method

Abstract

The invention discloses a swimming pool anti-drowning early warning system and method based on millimeter wave radar and vision fusion, and relates to the technical field of safety early warning. The system comprises a radar sensing unit, a visual identification unit and a central processing module. The radar sensing unit obtains basic motion parameters, micro-motion characteristic parameters and vital sign parameters of a target in a swimming pool in a non-contact mode through a millimeter wave radar, and abnormal analysis is carried out in combination with short-time and medium-time sliding windows to generate preliminary early warning. The visual identification unit is started after the radar sensing unit triggers the preliminary early warning, and target classification, attitude estimation and behavior analysis are carried out through a camera; the central processing module fuses radar and visual data, carries out multi-dimensional risk assessment and outputs graded early warning signals; according to the invention, all-weather, high-sensitivity and low-false-alarm drowning early warning is realized, graded response and intelligent rescue scheduling are supported, and the system is suitable for safety protection of deepwater places such as swimming pools and the like.

Description

Technical Field

[0001] This invention relates to the field of safety early warning technology, specifically to a swimming pool drowning prevention early warning system and method based on millimeter-wave radar and visual fusion. Background Technology

[0002] Drowning accidents in deep-water areas such as swimming pools and water parks pose a serious threat to public safety. Traditional safety measures rely primarily on the visual monitoring of lifeguards, with lifeguards observing the water from elevated towers erected on the shore. This method is highly dependent on individual focus and experience, and is prone to oversights when lifeguards are fatigued, have obstructed views, or are monitoring multiple people simultaneously, failing to provide continuous and reliable risk warnings. To improve the automation and intelligence of safety measures, the industry has developed various sensor-based solutions: deploying pressure sensor arrays on the pool walls or bottom, which detect changes in water pressure caused by a person sinking to the bottom. However, this method has a fatal flaw: it only triggers an alarm when the drowning person is completely submerged and stationary, far beyond the golden time for rescue; and deploying infrared transmitters and receivers around the pool to create an invisible defense, triggering an alarm if the beam is blocked by an intruder. Its essence is an intrusive identification and detection alarm. It can only determine whether someone has entered the water, but it cannot distinguish whether the person is a normal swimmer who actively jumps in or a person in distress. In actual use, the false alarm rate is extremely high. Pure visual detection has significant performance degradation in environments such as strong backlight, low light at night, water surface reflection, fog, and crowd obstruction. It also has limited ability to identify accidents such as silent drowning without obvious signs of struggle. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a swimming pool drowning prevention early warning system and method based on millimeter-wave radar and visual fusion, which solves the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a swimming pool drowning prevention early warning system based on millimeter-wave radar and visual fusion, comprising: The radar sensing unit is used to acquire basic motion parameters, micro-motion characteristic parameters and vital sign parameters of targets in the pool in real time through millimeter-wave radar, and to perform anomaly analysis to output preliminary warnings. The visual recognition unit is used to initiate linked monitoring after receiving an initial warning. It uses a camera to perform category recognition, posture analysis, and behavior analysis on the target and outputs data packets. The central processing module makes drowning prevention decisions based on the data packets extracted from the linkage monitoring.

[0005] Furthermore, the initial warning trigger includes: A short-term window and a medium-term window are set up. The short-term window is used to capture sudden behavioral abnormalities, and the medium-term window is used to capture continuous physiological deterioration trends. A structured data packet is output for each tracked target, which includes: target ID, basic motion parameters, micro-motion characteristic parameters, and vital sign parameters. The basic motion parameters include the world coordinates, instantaneous velocity, and motion trajectory of the target ID. The micro-motion characteristic parameters include the frequency and amplitude of limb micro-movements and the frequency of chest respiratory movements. The vital sign parameters include respiratory rate and heart rate. A time-series database is maintained for each target, storing all data packets of the most recent 60 seconds in strict timestamp order as the raw data pool for sliding window operations. When the accumulated data of a target ID in the raw data pool reaches the duration of the medium-term window, window analysis is triggered to obtain motion abnormality score, micro-motion abnormality score, and vital sign abnormality score, and analysis is performed based on these to determine whether to trigger a preliminary warning. There are preset thresholds for motion abnormality, micro-motion abnormality, and vital sign abnormality. When the motion abnormality score is greater than the motion abnormality threshold, a motion abnormality warning is generated; when the micro-motion abnormality score is greater than the micro-motion abnormality threshold, a micro-motion abnormality warning is generated; when the vital sign abnormality score is greater than the vital sign abnormality threshold, a vital sign abnormality warning is generated. If any one or more of the conditions for motion abnormality warning, micro-motion abnormality warning, and vital sign abnormality warning are triggered, a preliminary warning is generated. The preliminary warning includes the warning target ID, the world coordinates of the warning target, and the warning type.

[0006] Furthermore, the motor abnormality score: Instantaneous velocity sequences within a short window are extracted, and a two-dimensional Cartesian coordinate system is constructed with time as the x-axis and velocity as the y-axis. The instantaneous velocities are input into the two-dimensional coordinate axes, and the slope of velocity change over time is obtained through linear fitting, i.e., the gradient of velocity change within the short window. This yields the gradients S1 and S2 for two consecutive short windows. A velocity drop exceeding 0.6 m / s within a single second within a short window is identified as a velocity sag. If any velocity loss occurs within two consecutive short windows, an anomalous velocity loss score f is directly assigned to both consecutive short windows. 1 is 0.9; at the same time, the gradient changes of two consecutive short-term windows are calculated using the functions sigmoid(-k*S1) and sigmoid(-k*S2), and the maximum value of the two is taken as the velocity gradient anomaly score f2 within the two consecutive short-term windows; where k is a scaling factor; the maximum value of f1 and f2 in the two consecutive short-term windows is selected as the motion anomaly score for this window analysis and denoted as F, i.e., F=max(f1, f2); if there is no sudden loss of velocity in the two consecutive short-term windows, then f2 is directly taken as the motion anomaly score for this window analysis.

[0007] Furthermore, the micromotion anomaly score: The time series of limb micro-movement amplitudes within a short time window is extracted, and the peak point of the amplitude is detected. The window is then dynamically divided into rising and decaying phases based on this peak. The average amplitude within each phase is calculated, and the limb micro-movement energy ratio is obtained by dividing the average amplitude of the decaying phase by the average amplitude of the rising phase. This yields the limb micro-movement energy ratios for two consecutive short time windows. The maximum value is selected as the limb micro-movement energy ratio for the two consecutive short time windows, and a piecewise linear function is used to normalize it to obtain the energy anomaly score. The piecewise linear function is defined as follows: a preset energy ratio safety interval is defined, and the setting of the energy ratio safety interval is based on the comparison between normal swimming activities and drowning. Statistical differentiation of the evolution pattern of limb micro-movement energy during struggling behavior: When the limb micro-movement energy ratio is greater than the upper limit of the safe energy ratio range, it indicates stable energy with no significant exhaustion characteristics, and the amplitude of the energy abnormality score f3 is zero; when the limb micro-movement energy ratio is less than the lower limit of the safe energy ratio range, it indicates very significant exhaustion characteristics, and the amplitude of the energy abnormality score f3 is one; otherwise, that is, when the limb micro-movement energy ratio is within the safe energy ratio range, the energy abnormality score f3 is calculated using the formula: f3 = (upper limit of the safe energy ratio range - limb micro-movement energy ratio) / (upper limit of the safe energy ratio range - lower limit of the safe energy ratio range). Extract the time series of chest respiratory motion frequency within the mid-time window, and calculate the mean and standard deviation of the series. Divide the standard deviation by the mean to obtain the coefficient of variation. Normalize the coefficient of variation using a saturation function to obtain the chest motion abnormality score. The saturation function is: Chest motion abnormality score = 1 - exp(-λ * coefficient of variation), where λ is the sensitivity coefficient, which is set to 2.0 in this application scenario. The saturation function maps the coefficient of variation with a range of [0, +∞) to [0, 1), and the growth rate decreases, which conforms to the perception characteristics of physiological signals. A preset frequency threshold is set. If the mean of the chest respiratory motion frequency time series within the mid-time window is less than the frequency threshold, a very high chest abnormality score f4 of 0.95 is directly assigned. Otherwise, if the mean of the chest respiratory motion frequency time series within the mid-time window is greater than or equal to the frequency threshold, the chest abnormality score f4 is equal to the chest motion abnormality score. A weighting factor is assigned to both the energy anomaly score and the thoracic motion stability, and the energy anomaly score and the thoracic motion stability are then weighted and fused together to obtain the micromotion anomaly score.

[0008] Furthermore, abnormal physical signs score: Extract the respiratory rate time series within the median time window, calculate the median of the time series, and define safe respiratory rate intervals for adults and adolescents respectively. When the median is within the safe respiratory rate interval, the respiratory abnormality score f5 for that median time window is assigned zero. When the median is less than the lower limit of the safe respiratory rate interval, the respiratory abnormality score f5 is calculated using the formula: f5 = sigmoid((lower limit of safe respiratory rate interval - median) * g1), where g1 is the sensitivity coefficient. When the median is greater than the upper limit of the safe respiratory rate interval, the respiratory abnormality score is calculated using the formula: f5 = sigmoid((median - upper limit of safe respiratory rate interval) * g2), where g2 is the sensitivity coefficient. Similarly, the heart rate time series within the median time window is extracted, a preset heart rate safety interval is established, and the median of the heart rate time series is calculated. When the median is within the preset heart rate safety interval, it indicates a normal heart rate, and the abnormal heart rate score f6 for that median time window is assigned zero. When the median is less than the lower limit of the heart rate safety interval, it may indicate bradycardia or circulatory failure. The abnormal heart rate score f6 is calculated using the formula: f6=sigmoid((lower limit of the heart rate safety interval - median)*g3), where g3 is a sensitivity coefficient; the lower the heart rate, the closer the score is to 1. When the median is greater than the upper limit of the heart rate safety interval, it may indicate panic or stress-induced tachycardia. The abnormal heart rate score f6 is calculated using the formula: f6=sigmoid((median - upper limit of the heart rate safety interval)*g4), where g4 is a sensitivity coefficient.

[0009] Furthermore, the data packet output process includes: S2-1, automatically adjusts the camera gimbal pointing and zoom level according to the world coordinates of the warning target; S2-2: Determine the category of the warning target based on the deep learning model. If the category is determined to be a person, then execute S2-3; otherwise, interrupt the warning process. S2-3 calculates the immersion duration and vertical duration of the warning target based on human key point detection; integrates the warning target ID, world coordinates, dynamic anomaly score, micro-motion anomaly score, vital sign anomaly score, immersion duration and vertical duration into a data packet as the final output of the linkage monitoring, and reports it to the central processing module in real time.

[0010] Furthermore, the categories of early warning targets include: 截取包含预警目标的感兴趣区域图像,并将其输入一个专为泳池环境优化的深度学习目标分类模型,深度学习目标分类模型输出预警目标的类别标签及其对应的置信度分数;类别主要包括:成人、儿童、宠物、漂浮物;预设置信阈值,当预警目标通过模型输出的置信度小于置信阈值时,则将包含预警目标的视频流发送至中央处理模块进行显示,以进行人工的二次验证,确定预警目标的类别; 当预警目标通过模型输出的置信度大于或等于置信阈值时,若预警目标的类别为非人时,则立即终止当前预警流程,将本次事件记录为雷达误报,并向中央处理模块发送一条“目标为非人实体,关于预警目标的预警流程结束”的反馈。

[0011] 进一步地,浸没时长与垂直时长的计算过程包括: 预警目标被确认为人,则启用高精度人体姿态估计模型定位人体的17个标准骨骼关键点,具体的高精度人体姿态估计模型输出的是关键点在二维图像平面上的像素坐标序列,由此构成预警目标实时数字化的姿态骨架图像;利用摄像头标定得到的内外参数,结合泳池三维模型中已知的水面平面方程,将姿态骨架图像中鼻对应的像素坐标反投影至世界坐标系,并计算其与水面的垂直距离;以视频帧为单位进行连续判断,当鼻关键点被判定为低于水面时开始计时,统计其连续低于水面的总帧数,并将其除以摄像头在联动监测模式下的当前帧率得到浸没时长; 在图像平面上,分别计算左、右髋关节关键点的中点,以及左、右肩关节关键点中点,连接这两个中点形成的矢量,代表了躯干的主轴线方向;计算此躯干主轴线与图像垂直轴之间的夹角,得到身体垂直角;预设垂直角阈值,当身体垂直角小于垂直角阈值时,则判定预警目标在当前帧处于身体垂直姿态;以视频帧为单位进行连续判断,当身体垂直角被判定为小于阈值时开始计时,统计其连续处于身体垂直姿态的总帧数,并将其除以摄像头在联动监测模式下的当前帧率,得到垂直时长。

[0012] 进一步地,中央处理模块基于联动监测提取的数据包作出防溺水决策: Intercept the image of the region of interest containing the warning target and input it into a deep learning object classification model optimized for the pool environment. The deep learning object classification model outputs the class label of the warning target and its corresponding confidence score; the classes mainly include: adults, children, pets, floating objects; a confidence threshold is preset. When the confidence of the warning target output by the model is less than the confidence threshold, the video stream containing the warning target is sent to the central processing module for display for manual secondary verification to determine the class of the warning target. When the confidence of the warning target output by the model is greater than or equal to the confidence threshold, if the class of the warning target is non-human, the current warning process is immediately terminated, this event is recorded as a radar false alarm, and a feedback of "The target is a non-human entity, and the warning process for the warning target ends" is sent to the central processing module.

[0011] Furthermore, the calculation process of the immersion duration and the vertical duration includes: <000,041> If the warning target is confirmed as a person, a high-precision human pose estimation model is enabled to locate 17 standard skeletal key points of the human body. Specifically, the high-precision human pose estimation model outputs a sequence of pixel coordinates of the key points on the two-dimensional image plane, thus forming a real-time digital pose skeleton image of the warning target; using the internal and external parameters obtained by camera calibration and combining the known water surface plane equation in the pool three-dimensional model, the pixel coordinates corresponding to the nose in the pose skeleton image are back-projected into the world coordinate system, and its vertical distance from the water surface is calculated; continuous judgment is made in units of video frames. When the nose key point is determined to be below the water surface, timing starts, the total number of consecutive frames below the water surface is counted, and it is divided by the current frame rate of the camera in the linkage monitoring mode to obtain the immersion duration. On the image plane, calculate the midpoints of the left and right hip joint key points and the midpoints of the left and right shoulder joint key points respectively. The vector formed by connecting these two midpoints represents the main axis direction of the torso; calculate the angle between this torso main axis and the image vertical axis to obtain the body vertical angle; preset a vertical angle threshold. When the body vertical angle is less than the vertical angle threshold, it is determined that the warning target is in a body vertical pose in the current frame; continuous judgment is made in units of video frames. When the body vertical angle is determined to be less than the threshold, timing starts, the total number of consecutive frames in the body vertical pose is counted, and it is divided by the current frame rate of the camera in the linkage monitoring mode to obtain the vertical duration.

[0012] Furthermore, the central processing module makes a drowning prevention decision based on the data packets extracted by the linkage monitoring: Data packets from the linked monitoring system are extracted. If the immersion time exceeds 10 seconds or the vital signs abnormality score is greater than 0.9, a drowning warning is directly triggered. Otherwise, the immersion time and vertical duration are normalized to obtain immersion and vertical scores. Then, a weighting factor is assigned to each of the motion abnormality score, micro-motion abnormality score, vital signs abnormality score, immersion score, and vertical score. Based on this, a weighted fusion calculation is performed to obtain a risk score. A preset risk range is established. When the risk score is greater than the upper limit of the risk range, a drowning warning is triggered; when the risk score is within the risk range, an intervention warning is triggered; and when the risk score is less than the lower limit of the risk range, an observation warning is triggered. The drowning warning process is as follows: A bright red flashing light and a high-decibel continuous alarm are triggered throughout the entire pool. The system automatically switches to emergency broadcast, broadcasting an emergency message: "Lifeguards, please immediately proceed to the [warning target location]!" The system automatically shuts down equipment that may hinder rescue efforts, such as wave generators and circulating water pumps. A rescue mission is sent to the apps of at least two rescuers closest to the warning target, and the target's real-time location is continuously pushed to ensure timely rescue efforts. The intervention and early warning process is as follows: trigger the slow flashing of the orange warning light in the pool area, accompanied by intermittent, medium-volume beeping; automatically play pre-set voice prompts in the pool area, such as "Please be careful in the deep water area"; send an intervention task to the app of at least one rescuer at a distance from the warning target, and continuously push the real-time location of the target to intervene in the warning target in a timely manner. The observation and early warning process is as follows: the camera closest to the early warning target is selected to continuously observe the target, and the trajectory of the target is sent to the central processing module for display. The trajectory is rendered in yellow and flashed. This continues until a drowning warning, an intervention warning, or an updated radar sensing unit no longer triggers an early warning.

[0013] A swimming pool drowning prevention early warning method based on millimeter-wave radar and vision fusion includes the following steps: Step 1: Use millimeter-wave radar to acquire basic motion parameters, micro-motion characteristic parameters, and vital sign parameters of targets in the swimming pool in real time, and perform anomaly analysis to output preliminary warnings; Step two: Upon receiving the initial warning, initiate joint monitoring, use cameras to perform category identification, posture analysis, and behavior analysis on the target, and output data packets; Step 3: Make drowning prevention decisions based on the data packets extracted from the joint monitoring.

[0014] The present invention has the following beneficial effects: 1. This invention employs millimeter-wave radar to achieve stable monitoring around the clock and in all weather conditions. Millimeter-wave radar is unaffected by visual interference such as light, water surface reflection, and fog, thus compensating for the performance shortcomings of pure vision systems in harsh environments. Through micro-Doppler spectrum and phase demodulation technology, it non-invasively extracts vital signs such as respiratory rate and heart rate, enabling early detection of physiological abnormalities and overcoming the limitations of traditional sensors that require contact or bottoming out for triggering. Combined with short-term and medium-term sliding windows, it simultaneously captures sudden movement abnormalities and continuous physiological deterioration trends, improving the ability to identify silent drowning and other behaviors without obvious struggling. 2. The system of this invention maintains a low frame rate inspection mode under normal circumstances and only starts when the radar triggers an initial warning to conduct joint monitoring and analysis, which greatly reduces the system power consumption and computing resource consumption, making it suitable for continuous operation 24 / 7; it quickly distinguishes between human and non-human targets through a deep learning classification model, effectively filtering false alarms from the radar for floating objects, pets, etc., and improving the system's specificity; based on a high-precision attitude estimation model, it extracts behavioral features such as immersion time and body vertical angle, providing intuitive and quantifiable visual evidence for drowning determination; it automatically controls the camera pan-tilt to lock onto the target according to the radar coordinates, with a short response time, ensuring that the target is always in the center of the frame and the image is clear, facilitating continuous analysis; 3. This invention integrates radar motion, micro-motion, and vital sign anomaly scores with visual immersion and vertical attitude scores in a weighted manner to achieve multi-dimensional and hierarchical risk assessment, balancing sensitivity and reliability. Based on the risk score, it classifies warnings into three levels: observation, intervention, and drowning, with corresponding audio-visual prompts and rescue dispatch strategies to achieve precise response and resource optimization. It establishes direct alarm conditions of immersion > 10 seconds or vital sign anomalies > 0.9, ensuring that the highest level of rescue can be initiated immediately without complex calculations in extreme situations. It achieves spatiotemporal alignment, information complementarity, and collaborative judgment between radar and visual data, forming a closed-loop warning process of radar initial screening, visual confirmation, and central decision-making. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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 drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a system block diagram of the present invention; Figure 2 This is a block diagram of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0018] Application Scenario: Addressing the shortcomings of existing swimming pool drowning prevention early warning systems in terms of timeliness (inability to provide early warnings), accuracy (high false alarm / missed alarm rates), and environmental robustness (inability to operate 24 / 7), this paper proposes a swimming pool drowning prevention early warning system based on millimeter-wave radar and visual fusion. First, millimeter-wave radar is deployed: a frequency-modulated continuous wave radar chip (such as the TI IWR6843) with an operating frequency band of 77-81 GHz is selected, with a wavelength of approximately 3.9 mm, which is extremely sensitive to minute movements. One radar is deployed at each of the two midpoints of the two long sides of a standard rectangular swimming pool, at a height of 4 meters above the ground. The radar's elevation angle is adjusted downwards by 25 degrees to ensure its main beam completely covers the water surface and penetrates to the pool bottom. The two radars are synchronized at the microsecond level via GPS or Ethernet. Joint spatial calibration is performed using corner reflectors with known coordinates (placed at the four corners and center of the pool) to align the polar coordinate systems of the two radars to a world coordinate system with the pool center as the origin. A 4K PTZ camera with 30x optical zoom and preset positions is used; the camera angle can be adjusted to ensure coverage of the entire pool area. The cameras are deployed at high positions at the four corners of the pool to ensure overlapping and complete coverage without blind spots. Zhang Zhengyou's calibration method is used to calibrate the camera's intrinsic parameters. A radar-vision joint calibration board is used to solve for the transformation matrix of each camera relative to the world coordinate system, achieving accurate mapping from radar coordinates to pixel coordinates. All radars and cameras are connected to a local area network via a gigabit industrial switch to ensure low-latency data transmission. In practical applications, cameras need to be equipped with infrared supplementary lights or low-light sensors to ensure imaging quality at night and in low light conditions.

[0019] Please see Figure 1 The present invention provides a technical solution: a swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion, comprising: a radar sensing unit, a visual recognition unit and a central processing module; The radar sensing unit, in real-time communication with millimeter-wave radar deployed at high elevations around the pool, continuously scans the pool area to acquire basic motion parameters, micro-motion characteristics, and vital signs of all moving targets within the pool; specifically: The basic motion parameters of a moving target include distance, velocity, azimuth, and trajectory; the extraction of basic motion parameters includes: The radar transmits linear frequency modulated continuous waves at a frame rate of 30 Hz and receives reflected echoes to obtain the original intermediate frequency analog signal. After analog-to-digital conversion, a discrete complex signal sequence S_{raw}(n, m) is obtained, where n is the range gate index and m is the pulse sequence index. A 256-point FFT is performed on each pulse sequence S_{raw}(n, m) to obtain a range image with a resolution of 0.1 meters. At this point, the target's range can be obtained. A 128-point FFT is performed on each range gate S_{raw}(n, m) to obtain a velocity image with a resolution of 0.05 m / s. At this point, the target's radial velocity can be obtained. The CFAR (Constant False Alarm Rate) detection algorithm is applied to the range-Doppler two-dimensional matrix to extract valid point targets exceeding the noise threshold. Combined with a radar angle estimation algorithm (such as DBF), each point target is converted from polar coordinates (range, velocity, azimuth, elevation) to a three-dimensional point cloud (x, y, z, v) in the world coordinate system. The Global Nearest Neighbor (GNN) algorithm is used to associate point clouds between adjacent frames: for each tracked target, an extended Kalman filter is used to predict and update its state vector [x, y, z, v]^T, thereby outputting a continuous and smooth motion trajectory; thus, the system outputs the basic motion parameters of each tracked target in real time: world coordinates (distance, orientation), three-dimensional velocity, and continuous motion trajectory. Based on micro-Doppler spectrum analysis and time-frequency processing technology of radar echo signals, the frequency components modulated by the subtle vibrations of the target limbs and chest cavity are separated and analyzed. This allows for the non-contact extraction of refined vital signs parameters characterizing the target's motion state. Specific micro-motion characteristic parameters include limb micro-motion frequency, limb micro-motion amplitude, and chest cavity respiratory motion frequency. Micro-motion characteristic parameter extraction includes: From the above tracking process, a specific target (e.g., ID=001) is locked. From the radar's raw complex signal S_{raw}, a signal segment S_{target}(t) originating from the target's spatial region is extracted using digital beamforming technology. A short-time Fourier transform is applied to S_{target}(t). The window length is set to 256 sampling points with 50% overlap. A time-spectrum graph is generated, with time on the horizontal axis and Doppler frequency (corresponding to velocity) on the vertical axis. Energy is represented by color intensity. Within the 2-10Hz band of the time-spectrum graph, frequency points with concentrated energy are identified. By calculating the standard deviation and peak frequency of the signal energy within this band, it is quantified as limb micro-movement amplitude (dimensionless ratio) and limb micro-movement frequency (Hz). During struggle, this band has high energy and a diffused spectrum. Within the 0.1-0.5Hz ultra-low frequency band of the time-spectrum graph, a stable, periodic spectral line is identified. Its dominant frequency is determined using a peak detection algorithm; this is the frequency of chest cavity breathing movements, representing the frequency of chest cavity undulation. By analyzing the periodic phase changes caused by physiological activities in the aforementioned echo signals, and based on phase demodulation and physiological signal separation algorithms, interference such as limb movements is filtered out from the complex mixed signals, accurately restoring the rhythmic parameters characterizing the target's vital state. Specifically, the obtained vital sign parameters include respiratory rate and heart rate. It should be noted that the thoracic respiratory movement frequency is observed based on the micro-Doppler effect and is used to identify behavioral patterns with or without respiratory movements; the respiratory rate is the respiratory physiological rhythm measured based on the phase interference principle. The extraction of vital sign parameters includes: From the target signal S_{target}(t)=I(t)+jQ(t), its instantaneous phase is calculated as φ(t)=arctan[Q(t) / I(t)]. Since the output range of the arctan function is [-π,π], phase unwrapping is required to obtain a continuous phase change curve Φ(t). Ideally, the displacement d(t)=λ*Φ(t) / (4π), where λ is the wavelength (3.9mm). d(t) contains displacements caused by breathing, heartbeat, and limb movement. Passing d(t) through a 0.1-0.5Hz 4th order Butterworth bandpass filter yields a pure breathing displacement waveform d_brea. First, pass d(t) through a bandpass filter of 0.8-3.0Hz, then use independent component analysis (ICA) to separate the heartbeat signal d_heart(t) from possible respiratory harmonics and limb micro-motion noise. Calculate the power spectral density of the heartbeat signal d_breath(t) within a 60-second time window, and find the main peak in the 0.1-0.5Hz range; its frequency is the respiratory rate (breaths / minute). Calculate the power spectral density of d_heart(t), and find the main peak in the 0.8-3.0Hz range; its frequency is the heart rate (breaths / minute). At this point, the system outputs the respiratory rate and heart rate of each target within the pool area in real time. A preliminary assessment is made based on the target's basic motion parameters, micro-motion characteristic parameters, and vital sign parameters, including: To ensure that the pool drowning prevention early warning system achieves the best balance between high sensitivity and high reliability, this embodiment abandons the traditional single-time-window analysis method and adopts a dual-layer sliding window fusion judgment. Because drowning is not an instantaneous event, but a temporal process involving rapid behavioral changes and continuous physiological deterioration, a single-time-length observation window cannot simultaneously and accurately capture both types of characteristics. Therefore, this solution sets a short-time window and a medium-time window, which are used in conjunction for judgment. In practical application, the short-time window duration is set to 10 seconds, with a step size of 5 seconds, serving as the system's transient response window. The system is designed to precisely match the typical duration (5-20 seconds) of ineffective limb struggle and the instantaneous nature of the sudden drop in speed. This aims to quickly capture sudden behavioral abnormalities at the onset of a crisis with high temporal resolution, providing valuable lead time for early warning. The mid-time window is set to 30 seconds with a step size of 10 seconds. As a trend diagnosis window, its length covers the critical physiological transition period (30-60 seconds) from the occurrence of the accident to the loss of function and sinking. It also ensures that it includes at least 2-3 complete respiratory cycles, thereby reliably assessing the slow trend of continuous deterioration or failure of vital signs such as respiratory rate and heart rate, and effectively filtering out transient interference. The radar sensing unit outputs a structured data packet for each tracked target at a fixed frequency of 30Hz. The data packet includes: target ID (each target in the pool is assigned a unique ID), world coordinates, instantaneous velocity, frequency and amplitude of limb micro-movements, frequency of chest breathing movements, respiratory rate and heart rate; and maintains a time-series database for each target, storing all data packets of the most recent 60 seconds in strict order of timestamps as the raw data pool for sliding window operations. When the accumulated data for a target ID in the raw data pool reaches the duration of the medium-term window (i.e., 30 seconds), window analysis is triggered, which specifically includes: S1-1, Window analysis of basic motion parameters: As can be seen from the above settings for short-term and medium-term windows, each time window analysis is triggered, it includes two consecutive short-term windows and one medium-term window. For basic motion parameters, we use short-term windows for analysis: extract the instantaneous velocity sequence within the short-term window, construct a two-dimensional rectangular coordinate system with time as the abscissa and velocity as the ordinate, and input the instantaneous velocity into the two-dimensional coordinate axis. Through linear fitting, we obtain the slope of velocity change with time, that is, the gradient of velocity change within the short-term window. From this, we can obtain the gradients S1 and S2 of change between two consecutive short-term windows. If the velocity drop exceeds 0.6 m / s within one second within the short-term window, a velocity drop is marked. In this embodiment, the continuous time is set to 1 second, and the drop threshold is 0.6 m / s, which can be adjusted according to the actual application scenario requirements. If two consecutive short-term windows contain more than 0.6 m / s, the velocity drop will be recorded as a sudden drop. If any velocity loss occurs, a velocity loss anomaly score f1 of 0.9 is directly assigned to the velocity loss anomaly score within two consecutive short-term windows. Simultaneously, the sigmoid(-k*S1) and sigmoid(-k*S2) functions are calculated on the gradient changes within the two consecutive short-term windows, and the maximum value of these two is taken as the velocity gradient anomaly score f2 within the two consecutive short-term windows; where k is a scaling factor. It should be noted that the sigmoid function maps the instantaneous velocity V to [0,1], and the faster the velocity gradient decreases, the closer the score is. The maximum value between f1 and f2 in the two consecutive short-term windows is selected as the motion anomaly score for this window analysis, denoted as F, i.e., F = max(f1, f2). If no velocity loss occurs in either of the two consecutive short-term windows, f2 is directly taken as the motion anomaly score F1 for this window analysis. S1-2, Window analysis of micro-motion characteristic parameters: For micro-motion characteristic parameters, a combined short-time window and a medium-time window are used for analysis: the time series of limb micro-motion amplitudes within the short-time window is extracted, the peak point of the amplitude is detected, and the window is dynamically divided into rising and decaying phases based on this; the average amplitude within the two phases is calculated, and the average amplitude of the decaying phase is divided by the average amplitude of the rising phase to obtain the limb micro-motion energy ratio; it should be noted that this ratio effectively quantifies the burst-exhaustion pattern of struggling behavior, and a ratio significantly less than 1 is a dangerous feature; thus, the limb micro-motion energy ratios for two consecutive short-time windows can be obtained, and the maximum value is selected as the limb micro-motion energy ratio within the two consecutive short-time windows. A piecewise linear function is used to normalize this ratio to obtain the energy anomaly score; the piecewise linear function is: An energy ratio safety range is defined, based on the statistical distinction between the evolution patterns of limb micro-movement energy during normal swimming and drowning struggle. When the limb micro-movement energy ratio is greater than the upper limit of the energy ratio safety range, it indicates stable energy with no significant signs of exhaustion, and the energy anomaly score f3 is zero. When the limb micro-movement energy ratio is less than the lower limit of the energy ratio safety range, it indicates very significant signs of exhaustion, and the energy anomaly score f3 is one. Otherwise, when the limb micro-movement energy ratio is within the energy ratio safety range, the energy anomaly score f3 is calculated using the formula: f3 = (upper limit of the energy ratio safety range - limb micro-movement energy ratio) / (upper limit of the energy ratio safety range - lower limit of the energy ratio safety range). The time series of thoracic respiratory motion frequencies within the mid-time window is extracted, and the mean and standard deviation of the series are calculated. The coefficient of variation (COP) is obtained by dividing the standard deviation by the mean. The COP is a dimensionless value representing the ratio of the standard deviation to the mean. A higher COP indicates greater relative fluctuation of the respiratory frequency around its mean, i.e., a more unstable and disordered rhythm. The COP is normalized using a saturation function to obtain the thoracic motion abnormality degree. The saturation function is: thoracic motion abnormality degree = 1 - exp(-λ * COP), where λ is the sensitivity coefficient, which is set to 2.0 in this application scenario. The saturation function maps the COP from the range [0, +∞) to [0, 1), and the growth rate decreases, which conforms to the perception characteristics of physiological signals. To distinguish between disordered and weak signals, a very low frequency threshold is preset, which is derived from the physiological characteristics of respiratory failure at the end of the drowning process. Considering both theoretical limits and engineering reliability, this threshold is far lower than the minimum physiological resting respiratory rate of healthy adults or children (usually >8 breaths / minute). It aims to capture extreme dangerous states where respiratory function is on the verge of or has completely stopped. During the physiological process of drowning, breathing may change from rapid and disordered to weak and eventually stop due to hypoxia. When the average respiratory rate estimated by the radar is consistently lower than this threshold, it indicates that spontaneous breathing may have basically disappeared and vital signs are at an extremely critical level. In this application scenario, it is set to 4 breaths / minute. If the average value of the time series of chest cavity respiratory movement frequency in the mid-time window is less than the frequency threshold, it is considered that breathing may have stopped, and a very high chest cavity abnormality score f4 of 0.95 is directly assigned; otherwise, that is, when the average value of the time series of chest cavity respiratory movement frequency in the mid-time window is greater than or equal to the frequency threshold, the chest cavity abnormality score f4 is equal to the degree of chest cavity movement abnormality. A weighting factor is assigned to the energy anomaly score f3 and the thoracic motion stability f4 respectively, and the energy anomaly score f3 and the thoracic motion stability f4 are weighted and fused to obtain the micromotion anomaly score, which is denoted as F2. In this application scenario, the weighting factors of the energy anomaly score f3 and the thoracic motion stability f4 are set to 0.6 and 0.4 respectively. They can be adjusted according to changes in the application scenario or actual needs to emphasize different aspects. S1-3, Window analysis of vital signs parameters: For vital signs parameters, a median time window is used for analysis: the respiratory rate time series within the median time window is extracted, and the median of the time series is calculated. It should be noted that the median is more resistant to interference than the mean. Based on medical knowledge and swimming pool exercise physiology, safe respiratory rate ranges are set for adults and adolescents respectively. In this embodiment, they are set to [8 breaths / min, 25 breaths / min]. This range excludes excessively low frequencies at rest and excessively high frequencies after strenuous exercise. For children's pools, the upper limit of the range needs to be appropriately increased. When the median is within the safe respiratory rate range, it indicates normal breathing, and the median time window is assigned a value. The abnormal breathing score f5 is zero. When the median is less than the lower limit of the safe breathing rate range, it may indicate respiratory depression or cessation. The abnormal breathing score f5 is calculated using the formula: f5 = sigmoid((lower limit of safe breathing rate range - median) * g1), where g1 is the sensitivity coefficient. The lower the breathing rate, the closer the score is to 1. When the median is greater than the upper limit of the safe breathing rate range, it may indicate panic hyperventilation. The abnormal breathing score f5 is calculated using the formula: f5 = sigmoid((median - upper limit of safe breathing rate range) * g2), where g2 is the sensitivity coefficient. The heart rate time series within the median time window is extracted, and a preset heart rate safety interval is established. In this embodiment, it is set to [50 bpm, 150 bpm]. The heart rate abnormality score is calculated using the same threshold scoring method as the respiratory abnormality score calculation. Specifically, the median of the heart rate time series is calculated. When the median is within the preset heart rate safety interval, it indicates a normal heart rate, and the heart rate abnormality score f6 for that median time window is assigned zero. When the median is less than the lower limit of the heart rate safety interval, it may indicate bradycardia or circulatory failure. The heart rate abnormality score f6 is determined by the formula: f6 = sigmoid((heart rate safety interval)). The heart rate abnormality score is calculated as (median - median) * g3, where g3 is the sensitivity coefficient. The lower the heart rate, the closer the score is to 1. When the median is greater than the upper limit of the safe heart rate zone, it may indicate panic or stress-induced tachycardia. The abnormal heart rate score f6 is calculated using the formula: f6 = sigmoid((median - upper limit of the safe heart rate zone) * g4), where g4 is the sensitivity coefficient. It should be noted that the safe heart rate zone [50 beats / min, 150 beats / min] covers the physiological heart rate range from resting state to high-intensity exercise. The heart rate during drowning and struggling may exceed this range due to extreme stress. It should be noted that the core purpose of setting the sensitivity coefficients g1, g2, g3, and g4 is to linearly map the absolute degree to which physiological indicators deviate from the safe range to the urgency of risk perception. The sensitivity coefficients act as slope regulators of the sigmoid function, determining the sensitivity of the score to changes in outliers. g1 and g3 (corresponding to bradyventricular and cardiac arrhythmias) are usually set to relatively high values ​​(e.g., 0.5-1.0) because low respiratory / heart rate is directly related to hypoxia and circulatory failure, which are more critical signals requiring a rapid system response. Even a small deviation below the threshold should result in a significant increase in the risk score. g2 and g4 (corresponding to tachyventricular and cardiac arrhythmias) are set to relatively low values ​​(e.g., 0.2-0.5) because exercise and panic can also cause temporary increases in physiological indicators. Higher coefficients help avoid false alarms caused by brief, strenuous activity, assigning high scores only to sustained, extreme hyperventilation or tachycardia. The precise values ​​of all coefficients must be jointly calibrated using baseline data from the target scenario (such as children's / adult swimming pools) and drowning simulation data to ensure high sensitivity to real dangers while minimizing disruptive alarms caused by normal physiological fluctuations. Weighting factors were assigned to the abnormal breathing score f5 and the abnormal heart rate score f6, and a weighted fusion calculation was performed accordingly to obtain the abnormal vital signs score, denoted as F3. The basis for the weighting is that, in the physiological process of drowning, abnormal breathing is usually an earlier, more direct, and more specific marker, so its weight should be higher than that of abnormal heart rate. In this application scenario, the weighting factors for the abnormal breathing score f5 and the abnormal heart rate score f6 are set to 0.7 and 0.3, respectively. These weights can be calibrated and adjusted based on the baseline physiological data of different pool user groups or actual verification results. S1-4. There are preset abnormal motion thresholds, abnormal micro-motion thresholds, and abnormal physical sign thresholds. The setting of the abnormal motion threshold is based on the statistical differentiation of the target macroscopic motion patterns (such as speed and trajectory) between normal swimming behavior and disabled drowning states. Its value is usually set between the upper limit of the normal speed fluctuation range and the lower limit of the dangerous stationary state. The setting of the abnormal micro-motion threshold is derived from the joint statistical analysis of two refined features, namely the limb micro-motion energy ratio and the respiratory rhythm variation coefficient. This threshold defines the quantitative boundary between normal regular limb activities, stable breathing and ineffective struggling, and respiratory disorders. Its value needs to ensure that it can be effectively triggered when typical burst-failure micro-motion patterns and respiratory rhythm instability occur, and it is the core discriminant basis for early warning of abnormal behavior. The setting of the abnormal physical sign threshold is directly based on the medically recognized safety ranges of vital signs (respiration and heart rate) and the drowning physiological deterioration model. This threshold calibrates the threshold for vital signs to enter the clear moderately severe abnormal or even failure range from acceptable physiological fluctuations. Its setting gives priority to the timely response to the most dangerous conditions such as respiratory depression and cardiac arrest, and it is the ultimate safety baseline for the system to judge physiological risks. All thresholds are calibrated through a large amount of scenario data to maximize the warning sensitivity under the premise of controlling false alarms. When the abnormal motion score F1 is greater than the abnormal motion threshold, it indicates that the overall motion ability of the target has been significantly lost or a dangerous arrest has occurred, and an abnormal motion warning is generated. When the abnormal micro-motion score F2 is greater than the abnormal micro-motion threshold, it indicates that the target shows strong ineffective struggling behavior or a serious disorder in the breathing pattern, and an abnormal micro-motion warning is generated. When the abnormal physical sign score F3 is greater than the abnormal physical sign threshold, it indicates that the respiratory and circulatory vital signs of the target have entered the clear moderately severe abnormal range, and an abnormal physical sign warning is generated. If any one or more of the abnormal motion warning, abnormal micro-motion warning, and abnormal physical sign warning conditions are triggered, it indicates that the target has reached a dangerous level in at least one key risk dimension, and a preliminary warning is generated. The preliminary warning includes the warning target ID, the world coordinates of the warning target, and the warning type. The millimeter-wave radar is not affected by visual interference such as light, water surface reflection, and fog, and can achieve stable monitoring all day and all weather, making up for the performance shortcoming of the pure vision system in harsh environments. Through micro-Doppler spectrum and phase demodulation technology, it non-invasively extracts vital signs such as respiratory rate and heart rate to achieve early detection of physiological abnormalities, breaking through the limitations of traditional sensors that require contact or bottom triggering. By combining short-term and medium-term sliding windows, it synchronously captures sudden abnormal motions and continuous physiological deterioration trends, improving the recognition ability for silent drowning and other behaviors without obvious struggling.

[0020] The visual recognition unit is connected to the cameras arranged by the poolside in real time. When receiving the preliminary warning signal, it will perform linkage monitoring; otherwise, the cameras will continuously maintain a low frame rate patrol mode to save resources. It includes: When the system does not receive an initial warning signal from the radar sensing unit, the visual recognition unit automatically operates in a low-power inspection mode. The specific implementation is as follows: All high-definition network cameras deployed around the pool are set to perform cyclic scanning from several fixed preset positions covering the entire pool area. The image acquisition frame rate of each camera is dynamically adjusted to a low level of 1 to 2 frames per second (1-2 fps). In this mode, the system performs lightweight real-time processing on the video stream, primarily running a background subtraction algorithm based on a Gaussian Mixture Model (GMM) combined with moving area detection, rather than calling complex deep neural networks. This mode has low video stream data transmission bandwidth requirements and low algorithm computational complexity. Its overall computational and communication resource consumption typically does not exceed 15% of the system's total available resources, ensuring the system can operate continuously 24 / 7 without incurring high energy and computational costs. Upon receiving an initial warning from the radar sensing module, the visual recognition unit's operating mode is immediately forced to switch to the linkage monitoring mode, the specific implementation of which is as follows: The system analyzes the initial warning and extracts the warning target ID (consistent with the radar-tracked target ID), world coordinates (its coordinate system has been calibrated and aligned with the vision system), and warning type. Based on the calibration parameters of all cameras (position, orientation, field of view) and the world coordinates of the warning target ID, it selects one or two optimal cameras in real time that can observe the warning target ID most clearly and without obstruction. In practical applications, the selection strategy comprehensively evaluates the expected image size, angle, and relationship with environmental obstructions of the target in the camera's field of view. Precise gimbal control commands are sent to the selected cameras. These commands, through a coordinate transformation algorithm, convert the world coordinates of the warning target into the required pitch, yaw, and optical zoom of the camera's gimbal to obtain the best analysis resolution. Upon receiving the gimbal control commands, the camera's built-in servo mechanism completes gimbal rotation, lens zoom, and autofocus within 300 milliseconds, quickly and stably locking the warning target in the center of the image. The system ensures through feedback that the physical size of the target in the image meets the requirements of subsequent analysis, typically set so that the image height of a human target is not less than one-third of the total image height. After the camera successfully locks onto the warning target, it performs deep analysis of the target, from semantic recognition to pose quantization: S2-1: Extract the region of interest (ROI) image containing the warning target and input it into a deep learning target classification model optimized for swimming pool environments. This model is based on the YOLOv8 architecture and trained and fine-tuned on a large-scale dataset of swimming pool scene images (containing adults, children, common pets, and various floating objects under different lighting and angles). The model outputs the category label of the warning target and its corresponding confidence score. The categories mainly include: adults, children, pets, and floating objects (floating objects are mainly toys or swimming aids). A pre-set confidence threshold is used, based on the principle of high certainty in a safety-first warning system, requiring high certainty in the automatic identification results. The principle is that this threshold is intended to strictly distinguish between the model's confident and uncertain judgment states; in this embodiment, it is set to 0.9; when the confidence level of the warning target output by the model is less than the confidence threshold, it means that the model cannot accurately identify the category of the warning target, and the video stream containing the warning target is sent to the central processing module for display to perform a second manual verification. If the verification result is non-human (pet or floating object), the current warning process is immediately terminated, the event is recorded as a radar false alarm, and a feedback message "The target is a non-human entity, the warning process for the warning target is terminated" is sent to the central processing module; if the verification result is human (i.e., adult or child), then step S2-2 is executed; When the confidence level of the warning target output by the model is greater than or equal to the confidence threshold, it means that the model has accurately identified the category of the warning target; if the category of the warning target is non-human, the current warning process is terminated immediately, the event is recorded as a radar false alarm, and a feedback message "The target is a non-human entity, the warning process for the warning target has ended" is sent to the central processing module; if the category of the warning target is human, step S2-2 is executed. S2-2, if the warning target is confirmed to be a human, the high-precision human pose estimation model is activated. In this embodiment, HRNet (High-Resolution Net) is used as the backbone network. The high-precision human pose estimation model can maintain high-resolution feature maps throughout the entire processing, thereby accurately locating 17 standard skeletal key points of the human body, including nose, eyes, ears, shoulders, elbows, wrists, hips, knees, and ankles. Specifically, the high-precision human pose estimation model outputs the pixel coordinate sequence of the key points on the two-dimensional image plane, thus forming a real-time digital pose skeleton image of the warning target. Using the intrinsic and extrinsic parameters obtained from camera calibration, combined with the known water surface plane equation in the three-dimensional model of the swimming pool, the pixel coordinates corresponding to the nose in the pose skeleton image are back-projected to the world coordinate system, and its vertical distance from the water surface is calculated. Continuous judgment is performed on a video frame basis. When the nose key point is determined to be below the water surface, timing begins. The total number of consecutive frames below the water surface is counted and divided by the current frame rate of the camera in the linkage monitoring mode to obtain the immersion time. On the image plane, the midpoints of the key points of the left and right hip joints and the key points of the left and right shoulder joints are calculated respectively. The vector formed by connecting these two midpoints represents the direction of the main axis of the torso. The angle between this main axis of the torso and the vertical axis of the image (which is aligned with the direction of gravity after calibration) is calculated to obtain the body vertical angle. The smaller the body vertical angle, the closer the body posture is to vertical. Drowning victims often present an upright and struggling posture because they cannot float. A vertical angle threshold is preset. Based on the statistical analysis of a large number of normal swimming postures (such as the gliding phase of freestyle, breaststroke, and backstroke) and simulated postures of known drowning cases, the angle between the torso axis and the vertical direction is usually greater than 45 degrees during normal swimming. A typical upright struggling posture typically has an angle of less than 30 degrees. To achieve a balance between sensitivity and specificity, this embodiment sets the vertical angle threshold to 25 degrees. When the body's vertical angle is less than the threshold, it indicates that the target's body is in a dangerous posture that is extremely close to vertical, and the target is determined to be in a vertical posture in the current frame. Continuous judgment is performed on a video frame-by-frame basis. When the body's vertical angle is determined to be less than the threshold, timing begins, and the total number of consecutive frames in the vertical posture is counted. This count is then divided by the current frame rate of the camera in the linked monitoring mode to obtain the vertical duration. In practical applications, the vertical duration is a key quantitative indicator for assessing the persistence and severity of struggling behavior. In practical applications, a brief dip below the threshold may be a normal turn, while a vertical posture lasting more than 5 seconds is highly likely a sign of incapacitated struggling. S2-3 integrates the warning target ID, world coordinates, dynamic anomaly score F1, micro-motion anomaly score F2, vital sign anomaly score F3, immersion duration, and vertical duration into a structured data packet, which serves as the final output of the linkage monitoring and will be reported to the central processing module in real time. Normally maintaining a low frame rate inspection mode, it only activates high-precision joint monitoring and analysis when the radar triggers an initial warning, significantly reducing system power consumption and computing resource usage, making it suitable for continuous 24 / 7 operation; it quickly distinguishes between human and non-human targets through a deep learning classification model, effectively filtering false alarms from the radar for floating objects, pets, etc., and improving system specificity; based on a high-precision attitude estimation model, it extracts behavioral features such as immersion duration and body vertical angle, providing intuitive and quantifiable visual evidence for drowning determination; it automatically controls the camera pan-tilt to lock onto the target based on radar coordinates, with a short response time, ensuring that the target is always in the center of the frame and the image is clear, facilitating continuous analysis.

[0021] The central processing module makes drowning prevention decisions based on data packets extracted from the linkage monitoring, including: Extracting the data packets from the linked monitoring system, if the immersion duration exceeds 10 seconds or the abnormal vital signs score F3 is greater than 0.9, a drowning warning is directly triggered. It should be noted that 10 seconds of continuous head immersion far exceeds the safe breath-holding limit, a clear indication of bodily incapacitation and inability to surface independently; an abnormal score F3 greater than 0.9 means that vital signs (respiration / heart rate) have severely deviated from the normal range, indicating impending failure. The fulfillment of either of these two independent conditions indicates that the risk has progressed to an emergency stage that allows no delay; therefore, the system bypasses the conventional fusion calculation process and directly triggers the highest-level response. The values ​​of 10 seconds and 0.9 can be calibrated and adjusted according to the actual conditions of different application scenarios (such as children's pools, professional deep-water pools); otherwise, the immersion duration and vertical duration are normalized to obtain the immersion score and vertical score. The normalization process is as follows: (1) Immersion fraction = min((immersion time / 5 seconds), 1); (2) Vertical fraction = min((vertical duration / 3 seconds), 1); Next, a weighting factor is assigned to each of the following scores: motion abnormality score, micromotion abnormality score, vital sign abnormality score, immersion score, and vertical score. Based on this weighting factor, the risk score is calculated by weighted fusion of these scores. The weighting factor assignment is based on the credibility and importance of the features. In this embodiment, the vital sign abnormality score and immersion score have the highest weight (e.g., 0.3 each), followed by the motion abnormality score and micromotion abnormality score (e.g., 0.15 each), and then the vertical score (e.g., 0.1). The sum of all weighting factors is 1. A preset risk range is established, and the risk range is set through training with historical data. The training aims to balance the system's sensitivity (avoiding missed reports) and specificity (reducing false alarms), achieving precise matching between risk classification and response actions. In this embodiment, the risk range is set to [0.7, 0.9]. When the risk score is greater than the upper limit of the risk range, it indicates that a drowning emergency has been confirmed, requiring immediate initiation of the highest-level rescue response, thus triggering a drowning warning. When the risk score is within the risk range, it indicates that the target may be undergoing a drowning process, requiring immediate intervention, thus triggering an intervention warning. When the risk score is less than the lower limit of the risk range, the target is in a potential risk or early abnormal state, requiring continuous close monitoring, thus triggering an observation warning. The drowning warning process is as follows: A bright red flashing light and a high-decibel continuous alarm are triggered throughout the entire pool. The system automatically switches to emergency broadcast, broadcasting an emergency message: "Lifeguards, please immediately proceed to the [warning target location]!" The system automatically shuts down equipment that may hinder rescue efforts, such as wave generators and circulating water pumps. A rescue mission is sent to the apps of at least two rescuers closest to the warning target, and the target's real-time location is continuously pushed to ensure timely rescue efforts. The intervention and early warning process is as follows: trigger the slow flashing of the orange warning light in the pool area, accompanied by intermittent, medium-volume beeping; automatically play pre-set voice prompts in the pool area, such as "Please be careful in the deep water area"; send an intervention task to the app of at least one rescuer at a distance from the warning target, and continuously push the real-time location of the target to intervene in the warning target in a timely manner. The observation and early warning implementation process is as follows: the camera closest to the early warning target is selected to continuously observe the early warning target, and at the same time the trajectory of the early warning target is sent to the central processing module for display, and the trajectory is displayed in yellow and flashing; until a drowning warning or intervention warning is issued or the updated radar sensing unit no longer triggers the early warning; By weighted and fused with radar motion, micro-motion, and vital sign anomaly scores, along with visual immersion and vertical attitude scores, a multi-dimensional and hierarchical risk assessment is achieved, balancing sensitivity and reliability. Based on the risk score, three levels of early warning are established: observation, intervention, and drowning, with corresponding audio-visual prompts and rescue dispatch strategies to achieve precise response and resource optimization. Direct alarm conditions are set for immersion > 10 seconds or abnormal vital signs > 0.9, ensuring that the highest level of rescue can be initiated immediately without complex calculations in extreme situations. The system achieves spatiotemporal alignment, information complementarity, and collaborative judgment between radar and visual data, forming a closed-loop early warning process of radar initial screening, visual confirmation, and central decision-making.

[0022] A swimming pool drowning prevention early warning method based on millimeter-wave radar and vision fusion includes the following steps: Step one: Real-time acquisition of basic motion parameters, micro-motion characteristic parameters, and vital sign parameters of targets within the swimming pool using millimeter-wave radar; anomaly analysis is then performed to generate an initial warning; this initial warning includes: A short-term window and a medium-term window are set up. The short-term window is used to capture sudden behavioral abnormalities, and the medium-term window is used to capture continuous physiological deterioration trends. A structured data packet is output for each tracked target, which includes: target ID, basic motion parameters, micro-motion characteristic parameters, and vital sign parameters. The basic motion parameters include the world coordinates, instantaneous velocity, and motion trajectory of the target ID. The micro-motion characteristic parameters include the frequency and amplitude of limb micro-movements and the frequency of chest respiratory movements. The vital sign parameters include respiratory rate and heart rate. A time-series database is maintained for each target, storing all data packets of the most recent 60 seconds in strict timestamp order as the raw data pool for sliding window operations. When the accumulated data of a target ID in the raw data pool reaches the duration of the medium-term window, window analysis is triggered to obtain motion abnormality score, micro-motion abnormality score, and vital sign abnormality score, and analysis is performed based on these to determine whether to trigger a preliminary warning. There are preset thresholds for motion abnormality, micro-motion abnormality, and vital sign abnormality. When the motion abnormality score is greater than the motion abnormality threshold, a motion abnormality warning is generated; when the micro-motion abnormality score is greater than the micro-motion abnormality threshold, a micro-motion abnormality warning is generated; when the vital sign abnormality score is greater than the vital sign abnormality threshold, a vital sign abnormality warning is generated. If any one or more of the conditions for motion abnormality warning, micro-motion abnormality warning, and vital sign abnormality warning are triggered, a preliminary warning is generated. The preliminary warning includes the warning target ID, the world coordinates of the warning target, and the warning type.

[0023] Step two: Upon receiving the initial warning, initiate coordinated monitoring, using cameras to perform target category identification, posture analysis, and behavior analysis, and output data packets; the data packet output process includes: S2-1, automatically adjusts the camera gimbal pointing and zoom level according to the world coordinates of the warning target; S2-2: Determine the category of the warning target based on the deep learning model. If the category is determined to be a person, then execute S2-3; otherwise, interrupt the warning process. S2-3, calculate the immersion duration and vertical duration of the warning target based on human key point detection; integrate the warning target ID, world coordinates, dynamic anomaly score, micro-motion anomaly score, vital sign anomaly score, immersion duration and vertical duration into a data packet as the final output of the linkage monitoring; Step 3: Make drowning prevention decisions based on the data packets extracted from the joint monitoring system, specifically including: Data packets from the linked monitoring system are extracted. If the immersion time exceeds 10 seconds or the vital signs abnormality score is greater than 0.9, a drowning warning is directly triggered. Otherwise, the immersion time and vertical duration are normalized to obtain immersion and vertical scores. Then, a weighting factor is assigned to each of the motion abnormality score, micro-motion abnormality score, vital signs abnormality score, immersion score, and vertical score. Based on this, a weighted fusion calculation is performed to obtain a risk score. A preset risk range is established. When the risk score is greater than the upper limit of the risk range, a drowning warning is triggered; when the risk score is within the risk range, an intervention warning is triggered; and when the risk score is less than the lower limit of the risk range, an observation warning is triggered. The drowning warning process is as follows: A bright red flashing light and a high-decibel continuous alarm are triggered throughout the entire pool. The system automatically switches to emergency broadcast, broadcasting an emergency message: "Lifeguards, please immediately proceed to the [warning target location]!" The system automatically shuts down equipment that may hinder rescue efforts, such as wave generators and circulating water pumps. A rescue mission is sent to the apps of at least two rescuers closest to the warning target, and the target's real-time location is continuously pushed to ensure timely rescue efforts. The intervention and early warning process is as follows: trigger the slow flashing of the orange warning light in the pool area, accompanied by intermittent, medium-volume beeping; automatically play pre-set voice prompts in the pool area, such as "Please be careful in the deep water area"; send an intervention task to the app of at least one rescuer at a distance from the warning target, and continuously push the real-time location of the target to intervene in the warning target in a timely manner. The observation and early warning implementation process is as follows: the camera closest to the early warning target is selected to continuously observe the early warning target, and the trajectory of the early warning target is displayed, and the trajectory is rendered in yellow and flashed; until a drowning warning or intervention warning is issued or the updated radar sensing unit no longer triggers the early warning.

[0024] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion, characterized in that, include: The radar sensing unit is used to acquire basic motion parameters, micro-motion characteristic parameters and vital sign parameters of targets in the pool in real time through millimeter-wave radar, and to perform anomaly analysis and output preliminary early warnings. The visual recognition unit is used to initiate linked monitoring after receiving an initial warning. It uses a camera to perform category recognition, posture analysis, and behavior analysis on the target and outputs data packets. The central processing module makes drowning prevention decisions based on the data packets extracted from the linkage monitoring.

2. The swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion according to claim 1, characterized in that, Preliminary warning triggers include: A short-term window and a medium-term window are set up. The short-term window is used to capture sudden behavioral abnormalities, and the medium-term window is used to capture continuous physiological deterioration trends. A structured data packet is output for each tracked target, which includes: target ID, basic motion parameters, micro-motion characteristic parameters, and vital sign parameters. The basic motion parameters include the world coordinates, instantaneous velocity, and motion trajectory of the target ID. The micro-motion characteristic parameters include the frequency and amplitude of limb micro-movements and the frequency of chest respiratory movements. The vital sign parameters include respiratory rate and heart rate. A time-series database is maintained for each target, storing all data packets of the most recent 60 seconds in strict timestamp order as the raw data pool for sliding window operations. When the accumulated data of a target ID in the raw data pool reaches the duration of the medium-term window, window analysis is triggered to obtain motion abnormality score, micro-motion abnormality score, and vital sign abnormality score, and analysis is performed based on these to determine whether to trigger a preliminary warning. There are preset thresholds for motion abnormality, micro-motion abnormality, and vital sign abnormality. When the motion abnormality score is greater than the motion abnormality threshold, a motion abnormality warning is generated; when the micro-motion abnormality score is greater than the micro-motion abnormality threshold, a micro-motion abnormality warning is generated; when the vital sign abnormality score is greater than the vital sign abnormality threshold, a vital sign abnormality warning is generated. If any one or more of the conditions for motion abnormality warning, micro-motion abnormality warning, and vital sign abnormality warning are triggered, a preliminary warning is generated. The preliminary warning includes the warning target ID, the world coordinates of the warning target, and the warning type.

3. The swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion according to claim 2, characterized in that, Motor abnormality score: Instantaneous velocity sequences within a short window are extracted, and a two-dimensional Cartesian coordinate system is constructed with time as the x-axis and velocity as the y-axis. The instantaneous velocities are input into the two-dimensional coordinate axes, and the slope of velocity change over time is obtained through linear fitting, i.e., the gradient of velocity change within the short window. This yields the gradients S1 and S2 for two consecutive short windows. A velocity drop exceeding 0.6 m / s within a single second within a short window is identified as a velocity sag. If any velocity loss occurs within two consecutive short windows, an anomalous velocity loss score f is directly assigned to both consecutive short windows. 1 is 0.9; at the same time, the gradient changes of two consecutive short-term windows are calculated using the functions sigmoid(-k*S1) and sigmoid(-k*S2), and the maximum value of the two is taken as the velocity gradient anomaly score f2 within the two consecutive short-term windows; where k is a scaling factor; the maximum value of f1 and f2 in the two consecutive short-term windows is selected as the motion anomaly score for this window analysis and denoted as F, i.e., F=max(f1, f2); if there is no sudden loss of velocity in the two consecutive short-term windows, then f2 is directly taken as the motion anomaly score for this window analysis.

4. The swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion according to claim 3, characterized in that, Micromotion anomaly score: The time series of limb micro-movement amplitudes within a short time window is extracted, and the peak point of the amplitude of the limb micro-movement amplitude time series is detected. The window is then dynamically divided into rising and decaying phases based on this. The average amplitude within the two phases is calculated, and the average value of the decaying phase is divided by the average value of the rising phase to obtain the limb micro-movement energy ratio. The limb micro-movement energy ratios for two consecutive short time windows are obtained from this, and the maximum value is selected as the limb micro-movement energy ratio within the two consecutive short time windows. The ratio is then normalized using a piecewise linear function to obtain the energy anomaly score. Extract the time series of chest respiratory motion frequency within the mid-time window, and calculate the mean and standard deviation of the series. Divide the standard deviation by the mean to obtain the coefficient of variation. Normalize the coefficient of variation using a saturation function to obtain the chest motion abnormality degree. A preset frequency threshold is set. If the mean of the chest respiratory motion frequency time series within the mid-time window is less than the frequency threshold, a very high chest abnormality score f4 of 0.95 is directly assigned. Otherwise, if the mean of the chest respiratory motion frequency time series within the mid-time window is greater than or equal to the frequency threshold, the chest abnormality score f4 is equal to the chest motion abnormality degree. A weighting factor is assigned to both the energy anomaly score and the thoracic motion stability, and the energy anomaly score and the thoracic motion stability are then weighted and fused together to obtain the micromotion anomaly score.

5. The swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion according to claim 4, characterized in that, Abnormal physical signs score: The respiratory rate time series within the median time window is extracted, and a threshold scoring method is used to analyze and obtain the respiratory abnormality score. Specifically, this includes: calculating the median of the respiratory rate time series; setting safe respiratory rate intervals for adults and adolescents respectively; when the median is within the safe respiratory rate interval, the respiratory abnormality score f5 for that median time window is assigned to zero; when the median is less than the lower limit of the safe respiratory rate interval, the respiratory abnormality score f5 is obtained by the formula: f5=sigmoid((lower limit of safe respiratory rate interval - median)×g1), where g1 is the sensitivity coefficient; when the median is greater than the upper limit of the safe respiratory rate interval, the respiratory abnormality score is obtained by the formula: f5=sigmoid((median - upper limit of safe respiratory rate interval)×g2), where g2 is the sensitivity coefficient. Heart rate time series within the time window is extracted and analyzed using a threshold scoring method to obtain abnormal heart rate scores; weight factors are assigned to abnormal respiratory scores and abnormal heart rate scores respectively, and weighted fusion calculation is performed accordingly to obtain abnormal vital signs scores.

6. The swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion according to claim 5, characterized in that, The data packet output process includes: S2-1, automatically adjusts the camera gimbal pointing and zoom level according to the world coordinates of the warning target; S2-2: Determine the category of the warning target based on the deep learning model. If the category is determined to be a person, then execute S2-3; otherwise, interrupt the warning process. S2-3 calculates the immersion duration and vertical duration of the warning target based on human key point detection; integrates the warning target ID, world coordinates, dynamic anomaly score, micro-motion anomaly score, vital sign anomaly score, immersion duration and vertical duration into a data packet as the final output of the linkage monitoring, and reports it to the central processing module in real time.

7. The swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion according to claim 6, characterized in that, The categories of early warning targets include: The image of the region of interest containing the warning target is extracted and input into a deep learning target classification model optimized for the swimming pool environment. The deep learning target classification model outputs the category label of the warning target and its corresponding confidence score. The categories include: adults, children, pets and floating objects. A confidence threshold is preset. When the confidence score of the warning target output by the model is less than the confidence threshold, the video stream containing the warning target is sent to the central processing module for display, so as to perform secondary verification by humans to determine the category of the warning target. When the confidence level of the warning target output by the model is greater than or equal to the confidence threshold, if the category of the warning target is non-human, the current warning process is immediately terminated, the event is recorded as a radar false alarm, and a feedback message is sent to the central processing module that the target is a non-human entity and the warning process for the warning target has ended.

8. The swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion according to claim 7, characterized in that, The calculation process for immersion duration and vertical duration includes: Once the target is identified as a human, a high-precision human pose estimation model is used to locate 17 standard skeletal key points of the human body. Specifically, the high-precision human pose estimation model outputs a sequence of pixel coordinates of the key points on a two-dimensional image plane, thus forming a real-time digital pose skeleton image of the target. Using the intrinsic and extrinsic parameters obtained from camera calibration, combined with the known water surface plane equation in the three-dimensional model of the swimming pool, the pixel coordinates corresponding to the nose in the pose skeleton image are back-projected to the world coordinate system, and its vertical distance from the water surface is calculated. Continuous judgment is performed on a video frame-by-frame basis. When the nose key point is determined to be below the water surface, timing begins. The total number of consecutive frames below the water surface is counted and divided by the current frame rate of the camera in the linkage monitoring mode to obtain the immersion duration. On the image plane, the midpoints of the key points of the left and right hip joints and the midpoints of the key points of the left and right shoulder joints are calculated respectively. The vector formed by connecting these two midpoints represents the direction of the main axis of the torso. The angle between this main axis of the torso and the vertical axis of the image is calculated to obtain the body vertical angle. A vertical angle threshold is preset. When the body vertical angle is less than the vertical angle threshold, it is determined that the warning target is in a body vertical posture in the current frame. The judgment is continuously performed in video frames. When the body vertical angle is determined to be less than the threshold, the timing starts. The total number of frames in which it is continuously in a body vertical posture is counted and divided by the current frame rate of the camera in the linkage monitoring mode to obtain the vertical duration.

9. The swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion according to claim 8, characterized in that, The central processing module makes drowning prevention decisions based on data packets extracted from the linkage monitoring: Extract the data packets from the linkage monitoring. If the immersion time is greater than 10 seconds or the abnormal vital signs score is greater than 0.9, a drowning warning is triggered directly. Otherwise, the immersion time and vertical duration are normalized to obtain the immersion score and vertical score. Then, a weight factor is assigned to each of the abnormal motion score, abnormal micro-motion score, abnormal vital signs score, immersion score, and vertical score. Based on this, the abnormal motion score, abnormal micro-motion score, abnormal vital signs score, immersion score, and vertical score are weighted and fused to obtain a risk score. A risk range is preset. When the risk score is greater than the upper limit of the risk range, a drowning warning is triggered. When the risk score is within the risk range, an intervention warning is triggered; When the risk score is less than the lower limit of the risk range, an observation warning is triggered; The drowning warning process is as follows: triggering a red flashing light and a high-decibel continuous alarm throughout the entire pool area, automatically switching to emergency broadcast, playing corresponding content, and automatically shutting down the corresponding equipment; sending a rescue mission to the apps of at least two rescuers closest to the warning target, continuously pushing the real-time location of the warning target, and launching a rescue operation for the warning target; The intervention and early warning process is as follows: trigger the slow flashing of the orange warning light in the pool area, accompanied by intermittent, medium-volume beeping; automatically play pre-set voice prompts in the pool area, such as "Please be careful in the deep water area"; send an intervention task to the app of at least one rescuer at a distance from the warning target, and continuously push the real-time location of the warning target to intervene in the warning target in a timely manner. The observation and early warning implementation process is as follows: the camera closest to the early warning target is selected to continuously observe the early warning target, and at the same time the trajectory of the early warning target is sent to the central processing module for display, and the trajectory is displayed in yellow and flashing; until a drowning warning or intervention warning is issued or the updated radar sensing unit no longer triggers the initial warning.

10. A swimming pool drowning prevention early warning method based on millimeter-wave radar and vision fusion, applied to the swimming pool drowning prevention early warning system based on millimeter-wave radar and vision fusion as described in any one of claims 1-9; characterized in that, Includes the following steps: Step 1: Use millimeter-wave radar to acquire basic motion parameters, micro-motion characteristic parameters, and vital sign parameters of targets in the swimming pool in real time, and perform anomaly analysis to output preliminary warnings; Step two: Upon receiving the initial warning, initiate joint monitoring, use cameras to perform category identification, posture analysis, and behavior analysis on the target, and output data packets; Step 3: Make drowning prevention decisions based on the data packets extracted from the joint monitoring.

Images