A method for detecting falls in restrooms

By using hierarchical processing of dynamic and micro-moving point clouds and adaptive robust decision-making, the problems of false alarms and false negatives in fall detection in the complex environment of restrooms are solved, achieving efficient and accurate fall detection.

CN122410474BActive Publication Date: 2026-11-10SHANGHAI SONGCHUNGUO HEALTH TECH CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202610873521.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-06-17
Publication Date
2026-11-10
Estimated Expiration
2046-06-17

AI Technical Summary

Technical Problem

Existing fall detection methods in restrooms based on millimeter-wave radar are prone to false alarms and false negatives in complex environments, making it difficult to simultaneously achieve low false alarms and low false negatives, especially in the presence of multipath interference, water flow interference, and diverse fall actions.

Method used

A point cloud imaging architecture with micro-motion enhancement is adopted. Through hierarchical processing of dynamic point cloud and micro-motion point cloud, combined with a multi-level adaptive robust decision architecture, candidate regions for non-standing states are screened out. The decision threshold is adjusted by N/M criterion and interference confidence to achieve accurate detection of fall behavior.

Benefits of technology

It can effectively penetrate glass partitions to obtain obscured human body signals, reduce the false alarm rate, reduce false alarms, improve the detection accuracy and computing efficiency in the whole restroom scenario, and enhance the reliability of the judgment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122410474B_ABST
    Figure CN122410474B_ABST
Patent Text Reader

Abstract

This invention relates to a method for detecting falls in restrooms, comprising the following steps: based on radar echo signals, calculating a dynamic point cloud for capturing significant movements and a micro-motion point cloud for capturing minute movements; uniformly layering the dynamic and micro-motion point clouds along the height direction; and determining whether a fall has occurred based on the changes in the distribution of the dynamic and micro-motion point clouds at different height layers over time. Preferably, during the determination, candidate interval detection for non-standing states is first performed to narrow down the search range, and then a robust decision is achieved through a confirmation strategy that combines suspected fall detection, confirmed fall detection, interference source identification, and sliding window fusion, thereby effectively addressing the complex situation of numerous interference sources in restroom fall detection. This invention is applicable to different bathroom environments and different groups of people, and the algorithm has low complexity, allowing it to be deployed on edge devices.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of radar signal processing and restroom behavior detection technology, and in particular to a method for detecting falls in restrooms. Background Technology

[0002] With the deepening of the aging population and the expansion of the single-person living group, falls have become the leading cause of injury-related death among the elderly. Bathrooms, due to their slippery floors and confined spaces, are the areas with the highest fall rates in homes and public spaces. Millimeter-wave radar, with its inherent characteristics of being non-contact, requiring no image acquisition, and resistant to water mist, dust, and light interference, has become the preferred technology for bathroom fall detection.

[0003] However, existing methods for detecting falls in restrooms based on millimeter-wave radar have the following problems:

[0004] (1) The dense reflective surfaces in the bathroom (such as tiled walls, mirrors, glass partitions, etc.) lead to strong multipath and occlusion, making point clouds prone to missing or false targets appear; shower mist and water flow will introduce dynamic scattering, interfering with human body echoes;

[0005] (2) Falling movements are diverse (such as sitting, half-lying, and fully lying down) and there are significant individual differences, which makes the traditional height-based methods have insufficient generalization ability and difficult to achieve both low false alarms and low false negatives.

[0006] CN116840836 A discloses a fall detection method, which acquires the target trajectory and the associated point cloud associated with the target trajectory. By using multiple consecutive frames of associated point cloud, the average height, the degree of height change, and the degree of position change of the highest point of the target are determined to determine the scene type of the target. After determining the scene type, the energy height distribution of the associated point cloud in each frame is calculated, and the target is determined to be in an active or inactive state based on the proportion of point cloud energy within the height range. Then, the fall is determined based on the state change.

[0007] However, this method mainly relies on the height energy distribution of a single associated point cloud to determine the state. When the human body tends to stop moving after a fall, when the human body is blocked by water flow or glass partitions, or when there are dynamic / multipath interferences such as water flow, mirrors, or glass partitions in the bathroom, the human body-related point cloud may decrease or the energy may drop. The judgment based on the overall energy height distribution is easily dominated by water flow or high-altitude scattering points, resulting in misjudgment or missed judgment.

[0008] In addition, there are also low-posture behaviors in the restroom such as squatting to wash, bending over to wash hair, and sitting on a low stool to do laundry, which are not falls. If the fall judgment is based solely on the transition between active and inactive states, there is a lack of comprehensive verification of the persistence of low-height states, changes in the point cloud structure between layers, the persistence of small human movements, and the confidence of interference sources. It is difficult to simultaneously ensure low false alarms and low false negatives in the complex scenario of the entire restroom. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a method for detecting falls in restrooms, which can effectively deal with the complex situation of numerous interference sources in restroom fall detection and improve the accuracy of restroom fall detection.

[0010] The technical solution adopted by this invention to solve its technical problem is: to provide a method for detecting falls in restrooms, comprising the following steps:

[0011] Based on radar echo signals, dynamic point clouds for capturing significant motions and micro-motion point clouds for capturing minute motions are calculated respectively.

[0012] Dynamic point clouds and micro-motion point clouds are uniformly layered along the height direction. Based on the changes in the distribution of dynamic point clouds and micro-motion point clouds at different height layers over time, it is determined whether there is a fall behavior.

[0013] Furthermore, based on the changes in the distribution of dynamic point clouds and micro-motion point clouds at different altitude layers over time, it is determined whether a fall has occurred, including the following steps:

[0014] The point cloud features of the dynamic point cloud and the micro-motion point cloud are statistically analyzed frame by frame to form a layered point cloud feature sequence. The point cloud features include the first point cloud feature representing the global posture of the human body, the second point cloud feature representing the distribution state of the point cloud at each height layer, and the third point cloud feature representing the changes in the point cloud structure between height layers.

[0015] Based on the hierarchical point cloud feature sequence, the time window of the human body in a non-standing state is selected to obtain several candidate intervals of non-standing state with a set duration.

[0016] For a set number of consecutive non-standing state candidate intervals, suspected fall behavior is detected one by one. When the number of suspected fall behaviors detected is greater than the first threshold, the hierarchical point cloud feature sequences corresponding to the set number of consecutive non-standing state candidate intervals are extracted and merged to obtain a high-dimensional point cloud feature sequence.

[0017] Determine whether a fall has occurred based on the feature sequence of the high-dimensional point cloud.

[0018] Preferably, the first threshold can be dynamically adjusted based on the interference confidence level, whereby the interference confidence level is used to characterize the probability that the point cloud comes from a non-human interference source.

[0019] Preferably, if a fall is determined to have occurred, the interference confidence level is recalculated within the time window corresponding to the fall.

[0020] When the confidence level of the new interference is greater than the third set value and the point cloud feature pattern of the corresponding time window matches the typical interference source, the fall behavior judgment result is cancelled.

[0021] When the interference confidence level is less than the fourth set value or the point cloud feature pattern of the corresponding time window does not match the typical interference source, the fall behavior judgment result is retained and the final fall event is output.

[0022] Beneficial effects

[0023] By adopting the above-described technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0024] This invention employs a micro-motion enhanced point cloud imaging architecture, which extends the signal observation window through coherent accumulation along the slow time dimension. This amplifies the energy of weak and slow-motion signals such as human breathing and micro-movements of limbs after a fall, avoiding the fundamental omission of micro-movements of the human body as static clutter in conventional signal processing. At the same time, it can effectively penetrate glass partitions to obtain obscured human body signals, solving the problem of missed detection in slow-motion falls and glass-obstructed scenarios.

[0025] This application unifies the dynamic point cloud and micro-movement point cloud into layers along the height direction, extracts three complementary features: overall point cloud features, layered point cloud features, and inter-layer distribution comparison features, and combines them with a multi-level adaptive robust decision architecture based on the three types of point cloud features. It can be adapted to the entire bathroom scenario such as shower area, toilet area, and wash area, while taking into account the requirements of low false alarms and low false alarms.

[0026] This application estimates the user's height by using micro-movement point cloud with strong anti-interference ability, and generates a unique height threshold adapted to the user. This avoids the interference of water flow dynamic points on height estimation, and directly excludes risk-free states such as standing and walking, thereby narrowing the detection range and improving computing efficiency.

[0027] This application addresses the shortcomings of existing schemes that use fixed thresholds and thus cannot balance false alarms and missed alarms by adopting a fall detection triggering logic based on the N / M criterion and adaptively adjusting the decision threshold based on interference confidence. In addition, by adding confirmation decision and interference review steps, it can further filter out false alarms caused by occasional signal fluctuations and interference, significantly improving the reliability of the decision. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0029] Figure 2 This is a flowchart illustrating an embodiment of the present invention;

[0030] Figure 3 This is a schematic diagram of a robust fall detection process based on the N / M criterion and interference adaptation according to an embodiment of the present invention. Detailed Implementation

[0031] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0032] The embodiments of the present invention relate to a method for detecting falls in restrooms, such as... Figure 1 The above includes the following steps:

[0033] Based on radar echo signals, dynamic point clouds for capturing significant motions and micro-motion point clouds for capturing minute motions are calculated respectively.

[0034] Dynamic point clouds and micro-motion point clouds are uniformly layered along the height direction. Based on the changes in the distribution of dynamic point clouds and micro-motion point clouds at different height layers over time, it is determined whether there is a fall behavior.

[0035] Imaging human point clouds using millimeter-wave radar can acquire information on the target's distance, azimuth, pitch, and velocity, accurately capturing human motion characteristics and providing high-precision data support for behavioral pattern analysis. This implementation uses a 3-transmit, 4-receive MIMO radar module for experimentation, but is not limited to this radar radio frequency module.

[0036] Radar continuously emits electromagnetic wave signals into space. These signals are scattered or reflected by objects and received by the radar receiver. After passing through a signal amplifier, mixer, and ADC sampling, a discrete echo signal containing information in the range, time, and angle dimensions is obtained. Dynamic point clouds and micro-motion point clouds can be calculated in the following ways:

[0037] After performing time-frequency transformations along the fast and slow time dimensions of the radar echo signal, the first Doppler image is obtained through static clutter suppression processing.

[0038] After coherently accumulating the radar echo signal in the slow time dimension, time-frequency transformation is performed along the fast and slow time dimensions of the radar echo signal, and then static clutter suppression processing is applied to obtain the second Doppler image.

[0039] The first Doppler image and the second Doppler image were converted into dynamic point cloud and micro-motion point cloud, respectively.

[0040] Specifically, the echo received by the radar module can be represented as... ,in Let be the slow time dimension, indicating the th A linear frequency modulated continuous wave signal For fast time dimension, it means the first... Each sampling point Let be the antenna dimension, and let represent the th . The received signals from each channel. (The signal is then analyzed along the echo signal.) By performing a Fast Fourier Transform on the fast and slow times, and then applying static clutter suppression, a dynamic range-Doppler image containing target information is obtained. .

[0041] After static clutter suppression, the fine details of the human body point cloud, such as the point cloud of the body while squatting on the toilet or the torso point cloud while showering, may be lost. Therefore, to obtain a richer human body point cloud, the echo signal... Coherent accumulation in the slow time dimension, i.e., extending the observation time of the echo signal in the slow time dimension and improving the Doppler resolution, can yield the signal. ,in This is the concatenated slow-time dimension index. After clutter suppression and fast Fourier transform, the micro-motion distance-Doppler image can be obtained. Meanwhile, the signal-to-noise ratio is enhanced after coherent accumulation, which is beneficial for observing human signals behind the glass partition.

[0042] Target detection was performed on both dynamic range-Doppler and micro-range-Doppler images, and the angles of the detected target cells were estimated. After coordinate transformation, a dynamic point cloud containing rich human physical information was obtained. and micro-movement point cloud ,in These represent the target's x-coordinate, y-coordinate, and vertical coordinate, respectively. and These represent the number of points in the dynamic point cloud and the number of points in the micro-motion point cloud, respectively. Target detection can employ the Constant False Alarm Rate (CFAR) algorithm, and angle estimation can employ the Digital Beamforming (DBF) algorithm, but is not limited to these two algorithms.

[0043] In the scenario of fall detection in the restroom, the two types of point clouds are complementary. For example, when the shower is turned on, the water flow will generate a large number of dynamic point clouds and show a wide Doppler distribution; while human body scattering is more of a micro-motion point cloud at most times. Especially when a person falls and is close to the ground, the overall movement tends to be still, but there may still be breathing and micro-movements, which makes the micro-motion point cloud continue to exist in the low height area.

[0044] Therefore, the presence of a fall can be determined by observing the changes in the distribution of dynamic and micro-motion point clouds at different altitudes over time. For example... Figure 2 As shown, the specific steps include:

[0045] The point cloud features of the dynamic point cloud and the micro-motion point cloud are statistically analyzed frame by frame to form a layered point cloud feature sequence. Among them, the point cloud features include the first point cloud feature (overall point cloud feature) representing the global posture of the human body, the second point cloud feature (layered point cloud feature) representing the distribution state of the point cloud at each height layer, and the third point cloud feature (interlayer distribution comparison feature) representing the structural changes of the point cloud between height layers.

[0046] Based on the hierarchical point cloud feature sequence, a time window of a set duration in which the human body is in a non-standing state is selected, and several candidate intervals of non-standing state are obtained.

[0047] Further fall detection is performed on candidate regions in non-standing states; specifically, a robust fall detection framework based on the N / M criterion can be adopted, which mainly includes the following steps:

[0048] For a set number (denoted as M) of consecutive non-standing state candidate intervals, suspected fall behavior is detected one by one. When the number of suspected fall behaviors detected is greater than the set threshold (denoted as N), the hierarchical point cloud feature sequences corresponding to the set number of consecutive non-standing state candidate intervals are extracted and merged to obtain a high-dimensional point cloud feature sequence.

[0049] Determine whether a fall has occurred based on the feature sequence of the high-dimensional point cloud.

[0050] The aforementioned feature extraction method, based on the Doppler characteristics of two types of point clouds and a highly hierarchical approach, enhances adaptability to different bathroom structures and movement patterns. Specifically, the point cloud can be categorized by height... Direction is divided into Each layer calculates the dynamic and micro-motion point cloud features of each layer and the whole in each frame, mainly focusing on the following three categories:

[0051] (1) Overall point cloud features: used to depict the overall posture and global motion state of the human body, reflecting information such as whether it is at a low altitude and whether the whole is stable. It mainly includes: number of point clouds, Doppler characteristics, overall centroid height, highest / lowest altitude, proportion of low altitude points and their stability.

[0052] (2) Layered point cloud features: used to characterize the differences in scattering distribution at different heights, mainly including the number and proportion of layered points, layer morphology / stability, etc.

[0053] (3) Interlayer distribution contrast features: used to further highlight the changes in the height point cloud structure, such as the ratio of points in the lower layer to the upper layer, the combination of the proportion of micro-motion in the lower layer and the proportion of dynamic in the upper layer, etc.

[0054] Based on the extracted hierarchical point cloud feature sequences, candidate regions for non-standing states (including sitting on the toilet, squatting, washing clothes, etc.) can be quickly filtered out to narrow down the search range for subsequent fall detection, thereby improving computational efficiency while reducing false alarms. The following methods can be used:

[0055] Human height is estimated based on the hierarchical point cloud feature sequence of the human body when walking and / or standing, and then a height threshold is set.

[0056] The overall centroid height is periodically detected according to the set duration. If the overall centroid height is lower than the height threshold, and the proportion of low-level point clouds is greater than the set value, and the duration of this state exceeds the set value, then the current detection period is regarded as a candidate interval for non-standing state.

[0057] Standing and walking are the initial, routine actions of a person entering a restroom, and their point cloud characteristics are clearly physically distinct from those of sitting, squatting, falling, or being disturbed by water flow. For example, point clouds exhibit continuous horizontal displacement during walking, and their Doppler characteristics conform to the micro-motion patterns of human walking. In the standing state, point clouds show no continuous displacement but maintain overall height stability. Furthermore, the proportion of micro-motion points in the lower near-ground layers is very low during standing or walking. Therefore, based on point cloud fragments from walking and standing, human height can be estimated, and height thresholds can be determined. This implementation uses only micro-motion point clouds to estimate height, thus eliminating the interference of water flow during showering.

[0058] Determine the height threshold Then, the time window can be set to... A non-standing candidate region detection is performed at set intervals. If the human body's center of mass height is lower than... Furthermore, the low-level point cloud accounts for a relatively high proportion, and this state lasts for more than [time period missing]. If so, the human body is considered to be in a non-standing state.

[0059] The restroom contains many actions that are easily confused with falls, such as sitting on a low stool to wash clothes, squatting to wash hair or feet, and these actions also exhibit low-height characteristics. At the same time, the restroom environment contains many disturbances, such as running water in the toilet, water sloshing in the bucket, periodically vibrating washing machines, and hanging towels, which can also be easily confused with falls in radar echoes.

[0060] Therefore, to reduce false alarms in complex restroom environments, this method employs the following approach: Figure 3The robust fall detection framework shown is as follows: First, candidate regions for non-standing states are continuously detected. Then, suspected fall detection is performed segment by segment in the candidate regions, and interference confidence is estimated to estimate the probability that the current point cloud target is the interference source. Second, when the number of suspected falls in M ​​consecutive candidate regions reaches N (i.e., the N / M criterion is satisfied), a more stringent confirmed fall detection is triggered. The threshold for confirmed fall detection will be adaptively adjusted according to the interference. Finally, the confirmed fall detection results need to be further used for interference source identification.

[0061] The specific steps are as follows:

[0062] (1) Continuously acquire M candidate intervals: continuously detect candidate intervals in non-standing states and determine whether M candidate intervals are detected consecutively. If so, proceed to the next step of suspected fall detection.

[0063] (2) Suspected fall detection: Input the feature sequence of each candidate interval into a neural network such as LSTM to perform suspected fall detection, and count the number of suspected falls in M ​​intervals.

[0064] (3) Interference confidence estimation: During the suspected fall detection stage, the confidence of interference sources is estimated simultaneously for the candidate interval to characterize the strength of non-human dynamic scattering interference within the current interval. The interference confidence can be expressed as: .

[0065] (4) Interference adaptive threshold adjustment: Compare the number of suspected falls with the threshold N. If the number of suspected falls is... If N is reached, a fall detection confirmation is triggered. Simultaneously, based on the interference confidence level... The fall detection threshold is adaptively adjusted. When the interference confidence level is high, the decision threshold in the confirmation stage is increased or the low height duration requirement is extended to suppress false triggers caused by interference. When the interference confidence level is low, the default threshold is used to ensure sensitivity to real falls.

[0066] (5) After triggering the fall detection confirmation, the layered point cloud feature sequences corresponding to the M candidate intervals are concatenated in chronological order to obtain a high-dimensional point cloud feature sequence. The high-dimensional point cloud feature sequence includes, but is not limited to, the overall centroid height sequence, the low-height point cloud proportion sequence, the high-low layer point cloud quantity ratio sequence, the continuous frame count or continuous proportion of low-layer micro-motion point clouds, and the variance or rate of change of the point cloud distribution at each height layer. These features are used to characterize whether the human body is in a low-height state, whether the low-height state is continuous and stable, and whether the low-height state is accompanied by human micro-motion signals. The high-dimensional point cloud feature sequence is input into an SVM, random forest, lightweight neural network, or other machine learning model, which outputs a fall detection result or a detection score based on the above features. When the model's detection score meets the confirmation threshold, a fall is confirmed; otherwise, monitoring continues.

[0067] (6) Interference source verification: Recalculate the confidence of the interference source (or the interference category judgment result) within the time window corresponding to the confirmed fall, and verify it in combination with its temporal consistency; when the interference confidence is high and its characteristic pattern is consistent with the typical interference source, cancel the confirmed fall event detection; when the interference confidence is low or the interference pattern consistency is not met, retain the confirmed fall result and output the final fall event.

Claims

1. A method for detecting falls in restrooms, characterized in that, Includes the following steps: Based on radar echo signals, dynamic point clouds for capturing significant motions and micro-motion point clouds for capturing minute motions are calculated respectively. Dynamic point clouds and micro-moving point clouds are uniformly layered along the height direction; The point cloud features of the dynamic point cloud and the micro-motion point cloud are statistically analyzed frame by frame to form a layered point cloud feature sequence. The point cloud features include the first point cloud feature representing the global posture of the human body, the second point cloud feature representing the distribution state of the point cloud at each height layer, and the third point cloud feature representing the changes in the point cloud structure between height layers. The second point cloud feature includes the number of point clouds at each height layer and their proportion. Based on the hierarchical point cloud feature sequence, the time window of the human body in a non-standing state is selected to obtain several candidate intervals of non-standing state with a set duration. For a predetermined number of consecutive non-standing candidate intervals, suspected fall behavior is detected one by one. When the number of detected suspected fall behaviors exceeds a first threshold, the hierarchical point cloud feature sequences corresponding to the predetermined number of consecutive non-standing candidate intervals are extracted and merged to obtain a high-dimensional point cloud feature sequence. The first threshold is dynamically adjusted based on the interference confidence, which represents the probability that the point cloud originates from a non-human interference source. When the interference confidence is greater than a first predetermined value, the first threshold is increased; when the interference confidence is less than a second predetermined value, the first threshold is set to the default value. The presence of a fall is determined based on the high-dimensional point cloud feature sequence. If a fall is determined to exist, the interference confidence is recalculated within the time window corresponding to the fall. When the confidence level of the new interference is greater than the third set value and the point cloud feature pattern of the corresponding time window matches the typical interference source, the fall behavior judgment result is cancelled. When the interference confidence level is less than the fourth set value or the point cloud feature pattern of the corresponding time window does not match the typical interference source, the fall behavior judgment result is retained and the final fall event is output.

2. The detection method according to claim 1, characterized in that, Suspected fall behavior and interference confidence are predicted by inputting the hierarchical point cloud feature sequence of candidate regions in non-standing states into a suspected fall detection model based on a deep learning algorithm.

3. The detection method according to claim 1, characterized in that, The first point cloud feature includes the overall centroid height; based on the hierarchical point cloud feature sequence, a time window of a set duration in which the human body is in a non-standing state is selected, including the following steps: Human height is estimated based on the hierarchical point cloud feature sequence of the human body when walking and / or standing, and then a second threshold is set. The overall centroid height is periodically detected according to a set duration. If the overall centroid height is lower than the second threshold, and the proportion of low-level point clouds is greater than the third threshold, and the duration of this state exceeds the fourth threshold, then the current detection period is taken as a candidate interval for a non-standing state.

4. The detection method according to claim 3, characterized in that, When the interference confidence level is greater than the first set value, the fourth threshold is increased.

5. The detection method according to claim 3, characterized in that, Human height is estimated by using the point cloud features of micro-moving point clouds in the layered point cloud feature sequence.

6. The detection method according to claim 1, characterized in that, Determining whether a fall has occurred based on the feature sequence of high-dimensional point clouds is achieved using machine learning algorithms.

7. The detection method according to claim 1, characterized in that, Calculating dynamic point clouds for capturing salient motion and micro-motion point clouds for capturing minute motion includes the following steps: After performing time-frequency transformations along the fast and slow time dimensions of the radar echo signal, the first Doppler image is obtained through static clutter suppression processing. After coherently accumulating the radar echo signal in the slow time dimension, time-frequency transformation is performed along the fast and slow time dimensions of the radar echo signal, and then static clutter suppression processing is applied to obtain the second Doppler image. The first Doppler image and the second Doppler image were converted into dynamic point cloud and micro-motion point cloud, respectively.

Citation Information

Patent Citations

  • Indoor tumble detection method and device based on millimeter wave radar

    CN114296076A

  • Fall detection method, device and equipment based on indoor radar and medium

    CN118112558A

  • Millimeter wave radar human body target behavior identification method and system based on point cloud target tracking and attitude estimation

    CN122194126A