Millimeter wave radar multi-person fall detection method and system based on dense point cloud
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU DUOPU SURVEY TECH CO LTD
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-12
AI Technical Summary
Existing millimeter-wave radar fall detection technology struggles to handle complex scenarios involving multi-target interaction, occlusion, and similar actions, resulting in a high false alarm rate and failing to meet the practical needs of scenarios such as home-based elderly care.
A dense point cloud-based approach is adopted, using the HDBSCAN density clustering algorithm to separate human targets, combined with Kalman filter for continuous tracking, extracting multi-dimensional feature vectors and performing time-series analysis, establishing a multi-feature fusion fall judgment logic, and eliminating noise and environmental interference.
It significantly improves the spatial resolution and perception continuity of the radar for moving targets, effectively distinguishes between falls and daily actions, reduces the false alarm rate, and improves the practicality and stability of the system.
Smart Images
Figure CN122200799A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent sensing and health monitoring technology, and in particular to a method and system for detecting multiple falls using millimeter-wave radar based on dense point clouds, which is especially suitable for the accurate identification of people's fall behavior in scenarios such as home-based elderly care and smart healthcare. Background Technology
[0002] With the increasing aging of the population, falls among the elderly have become a significant public health issue. Current fall detection technologies primarily rely on wearable devices or visual monitoring equipment. However, wearable devices suffer from poor battery life and inconvenience, while visual devices are limited by privacy risks and sensitivity to lighting conditions. Millimeter-wave radar, with its advantages of non-contact sensing, all-weather monitoring, and strong anti-interference capabilities, has gradually become a research hotspot. However, existing millimeter-wave radar fall detection technology still has significant shortcomings and struggles to meet the practical needs of complex scenarios.
[0003] Existing technologies do not adequately consider complex scenarios such as falls occurring during arbitrary movements, in any direction, or when multiple targets coexist. For example, in a home environment, elderly people may suddenly fall while walking, turning, or sitting up, but existing methods are mostly based on pre-defined motion models and cannot cover the diversity of fall behaviors. Furthermore, in scenarios with multiple people present, interactions such as collisions or occlusions between targets can easily cause detection algorithms to fail.
[0004] In real-world environments, various negative actions of a target (such as bending over, squatting, and sitting up) are highly similar to falling behaviors in terms of motion characteristics, making them easily misjudged as falls. For example, the change in the height of the human center of gravity when bending over to pick up an item is similar to the process of falling. Existing methods rely on single features such as height thresholds for judgment, making it difficult to distinguish between real falls and everyday actions. Furthermore, environmental noise from homes and pets further exacerbates the false alarm problem, leading to a decrease in the system's usability. Summary of the Invention
[0005] In view of this, the purpose of the present invention is to provide a method and system for detecting multiple human falls using millimeter-wave radar based on dense point clouds. This method does not require a preset number of targets and can adaptively separate multiple human targets.
[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a millimeter-wave radar method for detecting multiple falls based on dense point clouds, comprising the following steps: The radio frequency signals received by the radar are processed to generate a dense point cloud containing three-dimensional spatial position coordinates, Doppler velocity and energy information; The point cloud was segmented using the HDBSCAN density clustering algorithm to separate the human target point cloud clusters and remove noise points and environmental clutter. Continuous tracking of clustered multi-target objects is performed, and data association strategies are used to handle target intersection and occlusion issues. Extract the dynamic feature parameters of each tracked target and construct a multi-dimensional feature vector. The dynamic feature parameters include the projected area of the xy plane, the centroid height of the point cloud, the height of the head point cloud, and the vertical spatial span of the point cloud. A feature sliding window is established to store the multidimensional feature vector of each target in multiple consecutive frames. Based on the fall decision logic of multi-feature fusion, temporal analysis and pattern matching are performed on the feature vector of each target, and the fall detection result is output independently. The fall decision logic includes: judging whether the projected area shows a trend of first increasing and then decreasing within the feature sliding window, and judging whether the centroid height of the point cloud, the height of the head point cloud, and the vertical spatial span of the point cloud are continuously less than their respective length thresholds within the feature sliding window. If all conditions are met at the same time, it is determined that the target has fallen.
[0007] Furthermore, the step of generating a dense point cloud includes: Perform FFT on the fast time dimension to obtain the distance dimension FFT; The zero Doppler component is eliminated by subtracting adjacent pulses, thus preserving the moving target signal. Perform a two-dimensional fast Fourier transform to extract the target's distance and radial velocity information; A two-dimensional constant false alarm rate (CFAR) detection algorithm is used to screen target points and calculate their horizontal and vertical angles. The multi-frame moving point compensation algorithm is applied to extend the life cycle of points with non-zero velocity and energy exceeding the threshold in the previous several frames, and the intersection with the point cloud of the current frame is taken to form the point cloud set of the current frame.
[0008] Furthermore, the HDBSCAN density clustering algorithm constructs a minimum spanning tree and a hierarchical clustering tree by calculating the core distance and reach distance of each point, adaptively dividing multiple target point cloud clusters, and marking outliers for removal.
[0009] Furthermore, the continuous tracking step employs a Kalman filter, defining a state vector that includes the target position and velocity, and achieving target tracking through state prediction, covariance prediction, Kalman gain calculation, state update, and covariance update.
[0010] Furthermore, the data association strategy includes: Iterate through all active tracking targets and spatially match them with the point cloud clustered in the current frame; When a point is associated with multiple targets, the point cloud is assigned based on the consistency of the motion direction. If the directions are consistent, the point is assigned to the nearest target; if the directions are inconsistent, the point is assigned to the target with the same motion direction.
[0011] Furthermore, the dynamic feature parameter extraction includes: Calculate the projected area of the xy plane using the Shoelace formula; Calculate the point cloud height of the target head, the vertical spatial span of the point cloud, and the centroid height of the point cloud.
[0012] Furthermore, the extraction of dynamic feature parameters for each tracked target also includes the following steps: The projected area of the xy plane is obtained by calculating the convex hull of the target point cloud cluster in the xy plane and applying the Shoelace formula. According to the formula and Determine the maximum and minimum values of the point cloud cluster along the z-axis, where The z-axis coordinate of the i-th point in the point cloud cluster is used to obtain the height of the head point cloud. Vertical spatial span of point cloud ; Based on the formula for the centroid height of the target's point cloud Calculate the mean z-axis coordinate of the point cloud cluster to obtain the centroid height of the point cloud, where, For point cloud datasets The number of point clouds, For the current frame number The height of a point cloud.
[0013] Furthermore, it also includes steps to resolve the fall condition: After a fall event has been determined, it is determined whether the projected area shows a trend of first increasing and then decreasing in subsequent frames. At the same time, it is determined whether the centroid height of the point cloud, the head point cloud height, and the vertical spatial span of the point cloud are continuously greater than their respective recovery thresholds within the feature sliding window range. If all conditions are met at the same time, the fall state of the target is released.
[0014] Furthermore, it also includes a step for determining the effectiveness of the objective: Calculate the target's 3D projected area. If it remains below the human body size threshold, it is determined to be a non-human target, and the tracking trajectory is released.
[0015] The present invention provides a millimeter-wave radar multi-person fall detection system based on dense point clouds, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned method.
[0016] The beneficial effects of this invention are as follows: This invention provides a method and system for multi-person fall detection using millimeter-wave radar based on dense point clouds. This method significantly improves the spatial resolution and perception continuity of the radar for moving targets by generating high-density point clouds and combining them with a multi-frame moving point compensation algorithm. The HDBSCAN density clustering algorithm is used for point cloud segmentation, eliminating the need for a preset number of targets and adaptively separating multiple human targets while removing environmental noise and outliers. This method has good clustering capabilities for irregularly shaped targets, and is particularly suitable for recognizing non-rigid postures such as falls and sitting / lying positions. A Kalman filter framework is used to achieve continuous multi-target tracking, and a point cloud attribution mechanism based on motion direction consistency is introduced, effectively solving the ID switching and matching errors caused by target intersection and occlusion. This strategy maintains trajectory stability and continuity even in complex environments with frequent target interactions, significantly improving tracking reliability.
[0017] This method employs a multi-dimensional feature vector that integrates ground projection area, head height, vertical spatial span, and point cloud centroid height to comprehensively characterize human posture and motion state. Through a two-level fall detection logic combining conditional filtering and pattern matching, and by introducing a fall resolution mechanism, it achieves accurate identification of fall events and judgment of state recovery. This mechanism can effectively distinguish falls from easily confused movements such as bending over or squatting, significantly reducing the false alarm rate and improving the system's practicality.
[0018] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0019] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following drawings are provided for illustration.
[0020] Figure 1 This is the overall flowchart of the millimeter-wave radar multi-person fall detection method based on dense point clouds in this embodiment; Figure 2 This is a flowchart of the point cloud densification process in this embodiment; Figure 3 This is a schematic diagram of a dense point cloud of the human body in this embodiment; Figure 4 This is a flowchart of the multi-target tracking algorithm in this embodiment; Figure 5 This is a flowchart of the fall detection algorithm based on multi-dimensional feature fusion in this embodiment; Figure 6 This is a graph showing the changes in key characteristic parameters of the fall process in this embodiment; Figure 7This is a graph showing the changes in key characteristic parameters during the fall release process in this embodiment; Figure 8 This is a schematic diagram of the single-person fall detection results in a multi-person scenario in this embodiment; Figure 9 This is a schematic diagram of the multi-person fall detection results in a multi-person scenario in this embodiment. Detailed Implementation
[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.
[0022] Example 1 like Figure 1 As shown, the millimeter-wave radar multi-person fall detection method based on dense point clouds provided in this embodiment solves the problems of high false alarm rate and poor scene adaptability of existing technologies in complex scenarios by generating dense point clouds, dynamic joint decision-making, and multi-target tracking and positioning. Specifically, it includes the following steps: Step 1: Process the radio frequency signals received by the radar to obtain dense point cloud information, including three-dimensional spatial position coordinates, Doppler velocity, and energy; Step 2: The HDBSCAN density clustering algorithm is used to segment the high-density point cloud, effectively separating the human target point cloud clusters, while removing outlier noise points and environmental clutter. Step 3: Continuously track the clustered multi-targets and combine data association strategies to solve the problems of target intersection and occlusion, so as to ensure the stability and continuity of tracking; Step 4: For each point cloud sequence corresponding to the tracked target, extract its dynamic feature parameters, including the projected area of the xy plane, the centroid height of the point cloud, the height of the target head point cloud, and the vertical space span of the target point cloud, and construct a multi-dimensional feature vector representing the human posture and motion state. Step 5: Design a fall detection logic based on multi-feature fusion. Integrate the multi-dimensional feature vectors representing human posture and motion state in Step 4, perform temporal analysis and pattern matching on the feature vectors of each target, and independently output the fall detection results of each target. Furthermore, step one specifically includes the following sub-steps: Step 1.1: To effectively suppress steady-state interference, it is necessary to perform FFT on the fast time dimension to obtain the range dimension FFT, and then eliminate the zero Doppler component by subtracting adjacent pulses to retain the moving target signal.
[0023] Let the first Frame and the The echo signals are respectively and If so, the output will be: stationary targets Output ; the target of movement The signal is preserved.
[0024] Step 1.2: Perform a two-dimensional Fast Fourier Transform (FFT) on the signal to extract the target's range and radial velocity information in the range-Doppler domain. Then, a two-dimensional constant false alarm rate (CFAR) detection algorithm is used to filter targets on the range-Doppler image, effectively distinguishing real target points from noise background, and outputting the target points and their corresponding echo energy values. Based on this, a beamforming algorithm is applied to each detected target point, and its horizontal direction angle is calculated through array signal processing. and vertical direction angle The formula is as follows: in, For radar signal wavelength, The spacing between the antenna arrays. , This represents the number of array elements in the horizontal and vertical directions.
[0025] Based on this, the point cloud information is obtained: Step 1.3: To make the point cloud denser, a multi-frame moving point compensation algorithm is proposed. Before calculation... Points in a frame with non-zero velocity and energy greater than a set threshold have their lifecycle extended. The union of this lifecycle with the point cloud of the current frame is denoted as the point cloud set of the current frame, as shown in the following formula: Furthermore, in step two, the HDBSCAN density clustering algorithm is used to segment the high-density point cloud, effectively separating the human target point cloud clusters, while removing outlier noise points and environmental clutter. Specifically, by calculating the "core distance" and "reach distance" of each point, a minimum spanning tree is constructed and a hierarchical clustering tree is generated, thereby identifying potential target clusters at different density levels. Without pre-setting the number of targets, it can adaptively divide point cloud clusters corresponding to multiple targets and mark isolated noise points or environmental interference points as outliers for removal. Finally, each clustering result corresponds to an independent target.
[0026] Furthermore, in step three, the clustered multi-targets are continuously tracked, and a data association strategy is used to solve the problems of target intersection and occlusion, ensuring the stability and continuity of tracking. Specifically, the target velocity and position are input into the Kalman filter, and a state vector is initialized and defined, containing the target state. For a target in three-dimensional space, this vector contains position and velocity information, and the formula is: in These are the target's location coordinates, and The target's velocity components, and the initial estimation error covariance matrix. Process noise covariance matrix Measure the noise covariance matrix and the state transition matrix and observation matrix .
[0027] Using the state transition matrix Predict the target state at the next moment: ; here It is a control input matrix. This is the control vector. Then, predict the covariance: predict the error covariance at the current time step based on the error covariance matrix from the previous time step. ; Then, the Kalman gain is calculated: based on the prediction error covariance and the measurement noise covariance, the Kalman gain is calculated as follows: ; in, The observation matrix is used to transform the system's state space to the observation space. Then, the state estimate is updated, incorporating the actual measurements. Update the target state estimate with the predicted value. Next, update the covariance matrix, and update the error covariance matrix based on the new state estimate. ,in This represents the identity matrix. The final output of the Kalman filter is the corrected system state.
[0028] In multi-target tracking, to address point cloud allocation conflicts caused by occlusion or close-range interactions between targets, the system implements a refined point cloud attribution determination mechanism. Specifically, it traverses all currently active tracked targets and spatially matches each target with all clustered point clouds within the current frame. When a detection point is associated with multiple targets simultaneously, the system first analyzes the motion direction vectors of each target. If the motion directions of multiple targets are highly consistent, the point is assigned to the target with the closest spatial distance to improve spatial consistency; if the motion directions of the targets differ significantly, the point is preferentially assigned to the target with the same motion direction, thereby effectively suppressing target label confusion or mismatch caused by point cloud overlap. This strategy significantly enhances tracking robustness in complex scenarios such as target intersection and occlusion, avoiding target loss or ID swapping issues.
[0029] Furthermore, for each target in the associated point cloud, its spatial centroid coordinates are calculated, and multidimensional geometric features are extracted, including the target's projected area in the xy plane and its projected area in the xz plane. Projected area of the yz plane Calculate the three-dimensional projected area of the target. The area sequence is processed using the sliding window method. Dynamic analysis is used to determine whether the body size exceeds a preset threshold. If the three-dimensional projected area of the target Greater than If the area is consistently below a certain threshold, it is considered a valid human target and tracking is maintained; otherwise, if the area remains below a certain threshold... If the target is determined to be a non-human interference source such as a pet or a fan, the system will actively release the tracking trajectory, clear its historical trajectory information, and set the corresponding status to idle.
[0030] Furthermore, in step four, for each point cloud sequence corresponding to the tracked target, dynamic feature parameters are extracted, including the projected area in the xy plane, the centroid height of the point cloud, the height of the target head point cloud, and the vertical spatial span of the target point cloud, to construct a multi-dimensional feature vector representing the human posture and motion state. Specifically: The projected area in the xy-plane is calculated by determining the convex hull of the point cloud cluster containing the target in the xy-plane, and then using the Shoelace formula. The formula is expressed as follows: in The coordinates of the convex hull vertices are arranged in order. It is the calculated area.
[0031] Further, the goal The point cloud dataset to which the time point belongs is Each point The coordinates and velocity in three-dimensional space are: , Maximum height of point cloud , Minimum height .
[0032] From this, the point cloud height of the target's head can be determined. , Target point cloud Vertical spatial span at any moment .
[0033] The centroid height of the target's point cloud ; in for The number of point clouds, For the current frame number The height of a point cloud.
[0034] Further, in step five, the design is based on a multi-feature fusion fall detection logic. It integrates the multi-dimensional feature vectors representing human posture and motion state from step four, performs temporal analysis and pattern matching on the feature vectors of each target, and independently outputs the fall detection results for each target. Specifically: For each target, if it exists, a feature window is created to store the multidimensional feature data of the target for 50 frames.
[0035] When a fall occurs, the multidimensional features must meet the following condition: within these 50 frames, the area... The point cloud centroid height initially increases and then decreases across the entire observation window. Target head point cloud height and the vertical space span of the target point cloud The fall must remain below the threshold throughout the entire observation window. If the above conditions are met, it is confirmed as a genuine fall event.
[0036] When the target is able to move independently after falling, a fall release decision is set. The multi-dimensional features of the fall release decision must meet the following conditions: Within these 50 frames, area The point cloud centroid height initially increases and then decreases across the entire observation window. Target head point cloud height and the vertical space span of the target point cloud The value must remain above the threshold throughout the entire observation window. If the above conditions are met, the fall event is deactivated.
[0037] Example 2 This embodiment, in conjunction with the accompanying drawings and examples, provides a further detailed description of the method.
[0038] like Figure 1 As shown, this embodiment discloses a method for detecting multiple falls using millimeter-wave radar based on dense point clouds, including the following steps: Step 1: Process the radio frequency signals received by the radar to obtain dense point cloud information, including three-dimensional spatial position coordinates, Doppler velocity, and energy; Specifically, it includes the following sub-steps: Step 1.1: To effectively suppress steady-state interference, it is necessary to perform FFT on the fast time dimension to obtain the range dimension FFT, and then eliminate the zero Doppler component by subtracting adjacent pulses to retain the moving target signal.
[0039] Let the first Frame and the The echo signals are respectively and If so, the output will be: stationary targets Output ; the target of movement The signal is preserved.
[0040] Step 1.2: Perform a two-dimensional Fast Fourier Transform (FFT) on the signal to extract the target's range and radial velocity information in the range-Doppler domain. Then, a two-dimensional constant false alarm rate (CFAR) detection algorithm is used to filter targets on the range-Doppler image, effectively distinguishing real target points from noise background, and outputting the target points and their corresponding echo energy values. Based on this, a beamforming algorithm is applied to each detected target point, and its horizontal direction angle is calculated through array signal processing. and vertical direction angle The formula is as follows: in, For radar signal wavelength, The spacing between the antenna arrays. , This represents the number of array elements in the horizontal and vertical directions.
[0041] Based on this, the point cloud information is obtained: ; in, It represents a set of radar point clouds in one frame; This indicates the radial distance between the target point and the radar; Indicates the radial velocity of the target point; Indicates the horizontal azimuth of the target point; Indicates the pitch angle of the target point; This represents the echo energy at the target point; Step 1.3: As Figure 2 As shown, to make the point cloud denser, a multi-frame moving point compensation algorithm is proposed. Before calculation... Points in a frame whose velocity is not zero and whose energy is greater than a set threshold are allowed to continue their lifecycle. Their points are then joined with the points in the current frame, and this union is denoted as the point cloud set of the current frame, as shown in the following formula: ; in, This represents the current frame's point cloud set after densification; This represents the original point cloud set of the current frame; Indicates the first frame before the current frame. A collection of point clouds of frames; Represents a single point in a point cloud; Indicates the total number of frames compensated; like Figure 3 As shown, the resulting dense point cloud clearly displays the overall outline of the human body.
[0042] Step 2: The HDBSCAN density clustering algorithm is used to segment the high-density point cloud, effectively separating the human target point cloud clusters, while removing outlier noise points and environmental clutter. Specifically, by calculating the "core distance" and "reach distance" of each point, a minimum spanning tree is constructed and a hierarchical clustering tree is generated. This identifies potential target clusters at different density levels without pre-setting the number of targets. It can adaptively divide the point cloud clusters corresponding to multiple targets and mark isolated noise points or environmental interference points as outliers for removal. Ultimately, each clustering result corresponds to an independent target.
[0043] Step 3: As Figure 4 As shown, the target velocity and position are input into the Kalman filter, and a state vector is defined to initialize and define a state vector containing the target state. For a target in three-dimensional space, this vector contains position and velocity information, and the formula is: in, These are the target's location coordinates, and The target's velocity components, and the initial estimation error covariance matrix. Process noise covariance matrix Measure the noise covariance matrix and the state transition matrix and observation matrix .
[0044] Using the state transition matrix Predict the target state at the next moment: in, It is a control input matrix. This is the control vector. Then, predict the covariance: predict the error covariance at the current time step based on the error covariance matrix from the previous time step. ; in, This represents the prior estimation error covariance matrix; Then calculate the Kalman gain: Calculate the Kalman gain based on the predicted error covariance and the measurement noise covariance. ; in, The observation matrix is used to transform the system's state space into the observation space.
[0045] Then update the state estimate, incorporating the actual measurements. Update the target state estimate with the predicted value: ; in, Indicates Kalman gain; This represents the posterior state estimate; This represents the prior state estimate; Next, update the covariance matrix, and update the error covariance matrix based on the new state estimate: ; in, Represents the identity matrix.
[0046] The final output of the Kalman filter is the corrected system state.
[0047] The system then iterates through all currently active tracking targets and performs spatial matching between each target and all clustered point clouds within the current frame. When a detection point is associated with multiple targets simultaneously, the system first analyzes the motion direction vectors of each target. If the motion directions of multiple targets are highly consistent, the point is assigned to the target with the closest spatial distance to improve spatial consistency; if the motion directions of the targets are significantly different, the point is preferentially assigned to the target with the same motion direction.
[0048] Furthermore, for each target in the associated point cloud, its spatial centroid coordinates are calculated, and multidimensional geometric features are extracted, including the target's projected area in the xy plane and its projected area in the xz plane. Projected area of the yz plane Calculate the three-dimensional projected area of the target. The area sequence is processed using the sliding window method. Dynamic analysis is used to determine whether the body size exceeds a preset threshold. If the three-dimensional projected area of the target Greater than If the area is consistently below a certain threshold, it is considered a valid human target and tracking is maintained; otherwise, if the area remains below a certain threshold... If the target is determined to be a non-human interference source such as a pet or a fan, the system will actively release the tracking trajectory, clear its historical trajectory information, and set the corresponding status to idle.
[0049] Step 4: For each point cloud sequence corresponding to the tracked target, extract its dynamic feature parameters, including the projected area of the xy plane, the centroid height of the point cloud, the height of the target head point cloud, and the vertical space span of the target point cloud to construct a multi-dimensional feature vector representing the human posture and motion state. The projected area in the xy-plane is calculated by determining the convex hull of the point cloud cluster containing the target in the xy-plane, and then using the Shoelace formula. The formula is expressed as follows: ; in, The coordinates of the convex hull vertices are arranged in order. It is the calculated area; n represents the number of vertices; Further, the goal The point cloud dataset to which the time point belongs is Each point The coordinates and velocity in three-dimensional space are: , Maximum height of point cloud , Minimum height .
[0050] From this, the point cloud height of the target's head can be determined. , Target point cloud Vertical spatial span at any moment .
[0051] The centroid height of the target's point cloud ; in, For the sake of gathering clouds The number of point clouds, For the current frame number The height of a point cloud.
[0052] Step 5: As Figure 5As shown, the fall detection algorithm based on multi-dimensional feature fusion in this embodiment is as follows: First, iterate through each target. If the target exists, create a feature sliding window and then wait to store the multidimensional feature data of the target for 50 frames.
[0053] like Figure 6 The image shows the changes in point cloud multidimensional features over time during a human fall in a normal posture. It is clearly visible that the height of the center of mass gradually decreases and eventually remains at a low height, while the area of the human body in the xy-plane shows a trend of first increasing and then decreasing.
[0054] Furthermore, when a fall occurs, the multidimensional features must meet the following condition: within these 50 frames, the area... The point cloud centroid height initially increases and then decreases across the entire observation window. Target head point cloud height and the vertical space span of the target point cloud The fall must remain below the threshold throughout the entire observation window. If the above conditions are met, it is confirmed as a genuine fall event.
[0055] like Figure 7 The diagram illustrates the changes in multidimensional features of the point cloud over time from the occurrence of a fall to the recovery of behavior. After recovery, the height of the human body generally increases; the point cloud area, however, undergoes a process of decreasing in size.
[0056] Furthermore, when the target is able to move independently after falling, a fall release decision is set. The multi-dimensional features of the fall release decision must meet the following conditions: within these 50 frames, the area... The point cloud centroid height initially increases and then decreases across the entire observation window. Target head point cloud height and the vertical space span of the target point cloud The value must remain above the threshold throughout the entire observation window. If the above conditions are met, the fall event is deactivated.
[0057] Test data and algorithm verification: This embodiment uses a 60G millimeter-wave radar, and some parameters are as follows: Depend on Figure 8 As shown, the scenario is a two-person scenario. One person falls, while the other person moves normally within the detection area. Two tracking trajectories appear in the figure, each corresponding to one of the two test subjects. The multiple targets do not interfere with each other, and the fall detection algorithm accurately identifies targets that have fallen and those that have not. Figure 9As shown, the scenario is a two-person scenario where two people fall. As shown in the figure, there are two tracking trajectories corresponding to the two testers respectively. There is no ID swapping or large trajectory offset, and multiple targets do not interfere with each other. At the same time, the fall detection algorithm can accurately identify the two targets that have fallen.
[0058] The above embodiments are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. It should be noted that, for those skilled in the art, various modifications, substitutions, or equivalent variations can be made to the specific embodiments described herein without departing from the essence and core spirit of the present invention. For example, adjustments to the signal processing flow, clustering algorithm parameters, feature extraction dimensions, or decision logic order should all be considered within the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of the technical solutions claimed by the present invention.
Claims
1. A method for detecting multiple falls using millimeter-wave radar based on dense point clouds, characterized in that, Includes the following steps: The radio frequency signals received by the radar are processed to generate a dense point cloud containing three-dimensional spatial position coordinates, Doppler velocity and energy information; The point cloud was segmented using the HDBSCAN density clustering algorithm to separate the human target point cloud clusters and remove noise points and environmental clutter. Continuous tracking of clustered multi-target objects is performed, and data association strategies are used to handle target intersection and occlusion issues. Extract the dynamic feature parameters of each tracked target and construct a multi-dimensional feature vector. The dynamic feature parameters include the projected area of the xy plane, the centroid height of the point cloud, the height of the head point cloud, and the vertical spatial span of the point cloud. A feature sliding window is established to store the multidimensional feature vector of each target in multiple consecutive frames. Based on the fall decision logic of multi-feature fusion, temporal analysis and pattern matching are performed on the feature vector of each target, and the fall detection result is output independently. The fall decision logic includes: judging whether the projected area shows a trend of first increasing and then decreasing within the feature sliding window, and judging whether the centroid height of the point cloud, the height of the head point cloud, and the vertical spatial span of the point cloud are continuously less than their respective length thresholds within the feature sliding window. If all conditions are met at the same time, it is determined that the target has fallen.
2. The method for detecting multiple falls using millimeter-wave radar based on dense point clouds as described in claim 1, characterized in that, The steps for generating dense point clouds include: Perform FFT on the fast time dimension to obtain the distance dimension FFT; The zero Doppler component is eliminated by subtracting adjacent pulses, thus preserving the moving target signal. Perform a two-dimensional fast Fourier transform to extract the target's distance and radial velocity information; A two-dimensional constant false alarm rate (CFAR) detection algorithm is used to screen target points and calculate their horizontal and vertical angles. The multi-frame moving point compensation algorithm is applied to extend the life cycle of points with non-zero velocity and energy exceeding the threshold in the previous several frames, and the intersection with the point cloud of the current frame is taken to form the point cloud set of the current frame.
3. The method for detecting multiple falls using millimeter-wave radar based on dense point clouds as described in claim 1, characterized in that, The HDBSCAN density clustering algorithm constructs a minimum spanning tree and a hierarchical clustering tree by calculating the core distance and reach distance of each point, adaptively dividing multiple target point cloud clusters, and marking outliers for removal.
4. The method for detecting multiple falls using millimeter-wave radar based on dense point clouds as described in claim 1, characterized in that, The continuous tracking step employs a Kalman filter, defining a state vector that includes the target position and velocity. Target tracking is achieved through state prediction, covariance prediction, Kalman gain calculation, state update, and covariance update.
5. The method for detecting multiple falls using millimeter-wave radar based on dense point clouds as described in claim 4, characterized in that, The data association strategy includes: Iterate through all active tracking targets and spatially match them with the point cloud clustered in the current frame; When a point is associated with multiple targets, the point cloud is assigned based on the consistency of the motion direction. If the directions are consistent, the point is assigned to the nearest target; if the directions are inconsistent, the point is assigned to the target with the same motion direction.
6. The method for detecting multiple falls using millimeter-wave radar based on dense point clouds as described in claim 1, characterized in that, The dynamic feature parameter extraction includes: Calculate the projected area of the xy plane using the Shoelace formula; Calculate the point cloud height of the target head, the vertical spatial span of the point cloud, and the centroid height of the point cloud.
7. The method for detecting multiple falls using millimeter-wave radar based on dense point clouds as described in claim 1, characterized in that, The extraction of dynamic feature parameters for each tracked target also includes the following steps: The projected area of the xy plane is obtained by calculating the convex hull of the target point cloud cluster in the xy plane and applying the Shoelace formula. According to the formula and Determine the maximum and minimum values of the point cloud cluster along the z-axis, where The z-axis coordinate of the i-th point in the point cloud cluster is used to obtain the height of the head point cloud. Vertical spatial span of point cloud ; Based on the formula for the centroid height of the target's point cloud Calculate the mean z-axis coordinate of the point cloud cluster to obtain the centroid height of the point cloud, where, For point cloud datasets The number of point clouds, For the current frame number The height of a point cloud.
8. The method for detecting multiple falls using millimeter-wave radar based on dense point clouds as described in claim 1, characterized in that, It also includes steps to resolve a fall: After a fall event has been determined, it is determined whether the projected area shows a trend of first increasing and then decreasing in subsequent frames. At the same time, it is determined whether the centroid height of the point cloud, the head point cloud height, and the vertical spatial span of the point cloud are continuously greater than their respective recovery thresholds within the feature sliding window range. If all conditions are met at the same time, the fall state of the target is released.
9. The method for detecting multiple falls using millimeter-wave radar based on dense point clouds as described in claim 1, characterized in that, It also includes the step of determining the effectiveness of the objective: Calculate the target's 3D projected area. If it remains below the human body size threshold, it is determined to be a non-human target, and the tracking trajectory is released.
10. A millimeter-wave radar multi-person fall detection system based on dense point clouds, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method described in any one of claims 1 to 9.