Drowning detection verification method and system based on human factor closed loop feedback

Abstract

The application discloses a drowning detection verification method and system based on human factor closed-loop feedback. The method comprises the following steps: detecting the behavior data of a monitored person through a sensor, and generating an initial risk signal when a feature meeting a preset abnormal condition is detected; after receiving the initial risk signal, issuing a tactile stimulation signal through a wearable device worn by the monitored person; within a preset time window after issuing the tactile stimulation signal, continuously collecting the behavior data of the monitored person as response observation data through the sensor; comparing the response observation data with the behavior data before the tactile stimulation signal is issued, and automatically judging whether the monitored person produces an effective behavior response; updating the risk level according to the judgment result and outputting corresponding alarm information. The application introduces a human as an information node into a detection closed loop, automatically verifies the initial detection result through tactile challenge and behavior response, does not require the monitored person to perform manual operation, significantly reduces the false alarm rate at a very low incremental cost, while maintaining a high detection rate for real drowning events and an acceptable response delay. The application can be applied to drowning detection in places such as swimming pools, water parks and beaches.

Description

Technical Field

[0001] This invention relates to the field of water safety monitoring technology, specifically to a drowning detection and verification method and system. Background Technology

[0002] Drowning is a major public safety issue worldwide. While public swimming pools, water parks, and beaches are typically staffed with lifeguards, the large number of people, water reflections, and obstructed visibility make it difficult for lifeguards to detect all drowning incidents in a timely manner. Therefore, technology-based drowning detection systems have become an important supplementary tool.

[0003] Currently, drowning detection technologies for swimming pools and natural water bodies mainly include the following categories: AI visual detection solutions, wearable device solutions, underwater sonar or acoustic detection solutions, and multi-sensor fusion solutions.

[0004] AI vision detection solutions deploy cameras above swimming pools and combine computer vision algorithms to analyze swimmers’ postures and movement trajectories in real time to identify drowning characteristics. However, the detection accuracy drops significantly when there is water reflection, crowds, or turbid water, and the false alarm rate is high. Some products require manual labeling to gradually train the model to reduce false alarms, which relies on manual intervention and has a long training cycle.

[0005] Wearable device solutions analyze swimmers' posture by using wristbands or smartwatches equipped with built-in sensors such as accelerometers and gyroscopes. Chinese patent CN115457734A discloses a drowning detection wristband that vibrates to alert the wearer upon detecting drowning; the alarm is deactivated if the wearer manually cancels the vibration, otherwise it automatically alarms. Chinese patent CN108597186A discloses a drowning alarm method based on an individualized behavioral model, and CN206684903U discloses a smart anti-drowning wristband that allows manual deactivation of the alarm via a button. The vibration confirmation in these solutions is essentially a manual deactivation mechanism, implicitly assuming that healthy individuals will deactivate the alarm promptly, while drowning victims will not due to inability to operate the device. However, healthy swimmers may be unable to deactivate the alarm promptly for various reasons, leading to escalating false alarms; furthermore, this mechanism is merely a binary judgment of deactivation or non-deactivation, failing to utilize the device's existing motion sensors to automatically analyze behavioral changes after vibration to update the risk assessment.

[0006] Underwater sonar or acoustic detection solutions identify drowning behavior by deploying an array of acoustic sensors on the bottom of the pool to monitor underwater acoustic anomalies. However, underwater sonar has limited accuracy and a high false alarm rate when used alone, as normal activities such as splashing and diving can also generate noise and acoustic signatures.

[0007] Multi-sensor fusion solutions combine two or more of the aforementioned sensors and reduce false alarm rates through cross-validation logic. For example, a high-level alarm is triggered only when both a camera and sonar detect an anomaly. However, since all sensors only observe and do not intervene, and fusion only occurs at the data level, when the signals from different sensors are contradictory or there is significant noise, the system cannot proactively acquire additional information to assist in decision-making and can only wait conservatively, resulting in either delayed alarms or allowing false alarms to occur.

[0008] Beyond drowning detection, similar verification mechanisms exist in other safety fields. In fall detection for the elderly, some devices employ a tiered response mechanism where a vibration alert is followed by manual cancellation by the user, similar to the drowning detection wristband in CN115457734A, both relying on the monitored individual's active manual operation. In the field of computer networks, the Challenge-Response protocol verifies authenticity by having the authenticator actively send a probe and the requester return a response; however, this core concept has not yet been applied in the field of physical security monitoring.

[0009] In summary, existing drowning detection technologies share a fundamental flaw: all systems are passive observation systems—they only observe and do not respond. Specifically, existing systems lack active verification capabilities. After detecting abnormal signals, they can only rely on passive observation from multiple sensors for cross-verification or wait for manual action from the monitored individual, failing to actively probe the monitored individual's state. This leads to a contradiction between false alarm and false negative rates: increasing the detection threshold to reduce false alarms results in a higher false negative rate, while lowering the threshold to reduce false negatives causes a surge in false alarms. This contradiction is difficult to fundamentally resolve within a purely passive observation framework. Existing wristbands' manual cancellation mechanisms rely on the monitored individual actively performing manual actions, increasing the user's burden and failing to distinguish between those unable to operate due to drowning and those who, while not drowning, fail to respond promptly, still leaving room for false alarms. More critically, the monitored individual's state of consciousness, motor ability, and behavioral responses directly determine whether they are truly in danger, yet current technologies have never incorporated human behavioral responses as an actively detectable information source into the detection loop. Summary of the Invention

[0010] In view of the shortcomings of existing detection systems, which are all passive observation systems, unable to actively verify the status of the monitored person, and difficult to balance false alarm rate and false negative rate, the purpose of this invention is to provide a drowning detection verification method and system that can reduce the false alarm rate while maintaining a high detection rate.

[0011] To achieve the above objectives, the present invention provides a drowning detection and verification method based on human-caused closed-loop feedback, comprising the following steps.

[0012] S1. Detect the behavior data of the monitored person through one or more sensors. When the behavior data contains features that meet preset abnormal conditions, generate an initial risk signal.

[0013] S2. In response to the initial risk signal, a tactile stimulation signal is emitted to the monitored person through the wearable device worn by the monitored person.

[0014] S3. Within a preset time window after the tactile stimulation signal is emitted, the behavior data of the monitored person is continuously collected by the sensor as response observation data.

[0015] S4. The system compares the response observation data with the behavioral data before the tactile stimulus signal is emitted, and automatically determines whether the monitored person produces an effective behavioral response without relying on the monitored person to perform manual operations; wherein, the effective behavioral response refers to the behavioral changes produced by the monitored person within the preset time window that indicate that they have consciousness and action capabilities.

[0016] S5. Update the risk level based on the judgment result of the effective behavior response, and output the alarm information corresponding to the updated risk level.

[0017] The present invention also provides a drowning detection and verification system based on human-caused closed-loop feedback, including a sensor module, a wearable device, a processing module and an alarm module, wherein the processing module is configured to perform the steps of the above method.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0019] Furthermore, the criteria for an effective behavioral response include at least one of the following: the monitored person's head position changes from submerged to surface; the movement characteristics change from an abnormal pattern to a normal swimming pattern; and the acceleration or angular velocity data of the wearable device shows a characteristic change that matches the preset responsive action.

[0020] Furthermore, determining whether the monitored person produces an effective behavioral response includes at least one of the following methods: comparing the behavioral feature vectors before and after the tactile stimulus signal is emitted, and determining an effective behavioral response when the change exceeds a preset threshold; identifying whether the monitored person exhibits a preset responsive action after the tactile stimulus signal is emitted through a pre-trained machine learning model; and analyzing whether the abnormal behavioral features corresponding to the initial risk signal disappear or weaken after the tactile stimulus signal is emitted.

[0021] Furthermore, the tactile stimulation signal includes at least one level of tactile stimulation. When no effective behavioral response is detected after the first level of tactile stimulation signal is emitted, a second level of tactile stimulation signal is emitted and a second behavioral response judgment is performed, wherein the intensity of the first level of tactile stimulation signal is lower than the intensity of the second level of tactile stimulation signal.

[0022] Furthermore, the intensity, vibration pattern, and duration of the tactile stimulation signal are configured based on at least one of the following: the level of the initial risk signal, the age or swimming ability level of the monitored individual, the current environmental conditions, and the monitored individual's historical challenge-response record.

[0023] Furthermore, the one or more sensors include at least one of an image acquisition device deployed above the water surface, an acoustic sensor deployed underwater, and a motion sensor built into the wearable device.

[0024] Furthermore, updating the risk level includes: lowering the risk level or lifting the warning when a valid behavioral response is detected; maintaining or upgrading the risk level when no valid behavioral response is detected; and maintaining the current risk level and triggering the re-emission of tactile stimulation signals when the response judgment result is uncertain.

[0025] Furthermore, the method also includes a closed-loop feedback step: feeding back the judgment result of the effective behavioral response to the initial detection step, which is used to dynamically adjust the threshold parameters, risk model weights or response criterion thresholds of the monitored individuals for subsequent detection, wherein the response criterion thresholds of the monitored individuals are adaptively adjusted based on the historical challenge-response records of the monitored individuals.

[0026] Furthermore, the preset time window is 2 to 10 seconds.

[0027] Furthermore, the tactile stimulation signal is also accompanied by light and / or sound signals emitted by the wearable device.

[0028] Furthermore, the wearable device in the system also includes motion sensors, including accelerometers and gyroscopes. The system also includes a multi-sensor fusion unit for time-synchronizing and spatially correlating data from different sensors to achieve cross-validation. The processing module further includes a challenge decision unit for determining parameters of the tactile stimulus signal based on at least one of the following: the level of the initial risk signal, the monitored person's information, and environmental conditions.

[0029] Compared with existing technologies, this invention has the following beneficial effects. First, by using tactile stimulation challenges and automatic behavioral response detection, this invention can distinguish between false alarm scenarios that appear to be drowning but in which the individual is actually capable of responding, and real dangerous scenarios where the individual is truly drowning and unable to respond. Assuming that the initial detection module has a certain false alarm rate, and that normal individuals have a high probability of producing detectable behavioral changes in response to tactile stimulation, the residual false alarm rate after human factor verification will be significantly lower than the initial false alarm rate, theoretically achieving an order-of-magnitude reduction. Second, unlike the manual cancellation mechanism of existing technologies such as CN115457734A, the response detection of this invention is entirely completed automatically by the system—the monitored individual does not need to press any buttons or operate the interface; their natural instinctive reactions (such as lifting their head to breathe or changing the direction of movement) can be captured by the sensor system and used as response signals. This reduces the burden on the user and eliminates the possibility of misjudgment caused by the inability to manually cancel the response. Furthermore, this invention feeds back the behavioral response judgment results to the initial detection module for dynamic adjustment of detection parameters and risk model weights, forming a closed-loop optimization. In particular, by learning from an individual's historical challenge-response records, it can adaptively adjust the individual's response criterion threshold, improving personalized detection accuracy. Moreover, the tactile stimulation module (vibration motor) and motion sensor (accelerometer / gyroscope) of this invention are standard components in existing wearable devices, and the response analysis algorithm is implemented in software, requiring no additional hardware investment. Under existing wristband and sensor system deployments, the incremental cost of this invention is close to zero. Simultaneously, the delay introduced by the verification phase of this invention accounts for a small percentage of the golden rescue time for drowning, and during the verification period, a level one warning (lifeguard wristband vibration reminder) has already been issued, and lifeguards have already begun to pay attention, preventing delays in rescue of real drowning incidents. Finally, the challenge-response verification mechanism of this invention is not limited to drowning detection but can be extended to various safety monitoring scenarios that require distinguishing between real and abnormal signals, such as fall detection. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope of protection of the claims of the present invention. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the architecture of a drowning detection and verification system based on human-factor closed-loop feedback provided in an embodiment of the present invention.

[0032] Figure 2 This is a flowchart illustrating the drowning detection verification method based on human-factor closed-loop feedback provided in an embodiment of the present invention.

[0033] Figure 3This is a timing diagram for closed-loop verification provided in an embodiment of the present invention.

[0034] Figure 4 This is a schematic diagram of a multi-level challenge decision-making process provided in an embodiment of the present invention.

[0035] Figure 5 This is a schematic diagram of the structure of a wearable device provided in an embodiment of the present invention. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0037] This invention provides a drowning detection and verification system based on human-factor closed-loop feedback, the overall architecture of which is as follows: Figure 1 As shown, it includes a sensor module (10), a wearable device (20), a processing module (30), and an alarm module (40).

[0038] The sensor module (10) includes a surface image acquisition device (11) and an underwater acoustic sensor (12). The surface image acquisition device (11) is deployed above the pool (ceiling or pole mounted, 3-5m high), using a 4K wide-angle camera (IP68 protection rating) to collect visual image data of the water surface and surrounding areas. 4-6 units are deployed in a standard 50m×25m pool to achieve full coverage. The underwater acoustic sensor (12) is deployed at the bottom of the pool, using a low-frequency acoustic wave sensor (spacing approximately 5m) to collect underwater acoustic data. 8-12 sensors are deployed in a standard pool. The data collected by the above sensors is transmitted to the processing module (30) via wired or wireless means.

[0039] like Figure 5As shown, the wearable device (20) includes a tactile stimulation module (21), a motion sensor (22), a wireless communication module (23), a water immersion sensor (24), a microcontroller (25), and a battery (26). The tactile stimulation module (21) is a vibration motor used to send tactile stimulation signals to the monitored person when it receives instructions from the processing module. The vibration motor is selected considering the reduced vibration perception underwater, and a low-frequency strong vibration motor is preferred. The motion sensor (22) includes a three-axis accelerometer and a three-axis gyroscope, used to collect the motion data (acceleration, angular velocity) of the monitored person and transmit it to the processing module as part of the behavioral data. The wireless communication module (23) uses the LoRa or Bluetooth Low Energy (BLE) communication protocol for bidirectional communication with the processing module, i.e., receiving challenge instructions and uploading motion sensor data. The water immersion sensor (24) is used to detect whether the device is submerged in water. The microcontroller (25) is used for signal processing and communication management on the device side. The battery (26) is preferably a rechargeable lithium battery with a battery life of not less than 8 hours. Wearable devices can take the form of wristbands, smartwatches, or devices integrated into swimming caps. The preferred wearing position is on the inside of the wrist to enhance tactile sensitivity.

[0040] The processing module (30) is deployed on an edge server at the swimming pool site. It includes an initial detection unit (31), a challenge decision unit (32), a response analysis unit (33), and a risk update unit (34). The initial detection unit (31) receives data from the sensor module and uses computer vision algorithms (such as YOLO object detection + DeepSORT multi-object tracking) or voiceprint analysis algorithms to detect whether there are features in the behavior data of the monitored person that meet preset abnormal conditions. The abnormal conditions include, but are not limited to: the head being submerged in the water for more than a preset time, the body sinking vertically, the limbs exhibiting an irregular struggling movement pattern, the movement trajectory suddenly stopping, and abnormal underwater voiceprints. When an abnormal feature is detected, an initial risk signal is generated, which includes the risk level and the monitored person's identifier.

[0041] The challenge decision unit (32) receives the initial risk signal and determines the level (light / strong), vibration pattern, duration, and observation window length of the tactile stimulus signal based on the initial risk level, the monitored person's information (age, swimming ability level, etc.), environmental conditions (population flow, lighting, water quality, etc.), and historical challenge-response records. It then outputs the challenge command to the wearable device (20). The determination of the observation window length also takes into account the data sampling requirements of the built-in motion sensor (22) of the wearable device (20) after tactile stimulation, to ensure that the response analysis unit (33) can acquire the refined change characteristics related to the behavioral response in the acceleration and angular velocity data.

[0042] The response analysis unit (33) is used to receive the behavior data continuously collected by the sensor module as response observation data within a preset time window (preferably 3-8 seconds) after the tactile stimulation signal is emitted, and compare the response observation data with the behavior data before the tactile stimulation signal is emitted to determine whether the monitored person has produced an effective behavior response. The specific judgment methods include the following three, which can be used individually or in combination. Method 1 is time sequence comparison, which compares the behavior feature vector in the first time window (e.g., the first 3 seconds) before the tactile stimulation signal is emitted with the behavior feature vector in the second time window (e.g., the last 3 seconds) after the emission, calculates the distance between the two feature vectors (e.g., Euclidean distance or cosine similarity), and determines that the behavior response is effective when the distance exceeds a preset threshold. The behavior feature vector includes, but is not limited to, the quantified values ​​of dimensions such as head height, body tilt angle, movement speed, movement direction, limb movement frequency, and underwater voiceprint features. Method 2 is specific action recognition. It uses a pre-trained machine learning model (such as a CNN or LSTM-based action classification model) to identify whether the monitored person exhibits a preset responsive action after the tactile stimulus signal is emitted. The responsive actions include, but are not limited to, raising the head to breathe, changing the swimming direction, raising the wrist to check the device, and starting to move from a static state. These actions indicate that the monitored person has consciousness and the ability to act. The machine learning model is trained using swimming video data labeled with responsive actions. Method 3 is signal persistence analysis. It analyzes whether the abnormal behavioral features corresponding to the initial risk signal disappear or significantly weaken after the tactile stimulus signal is emitted. For example, if the initial abnormal feature is that the head is submerged in the water, and the head floats out of the water after the tactile stimulus, then the abnormal feature disappears and is judged as a valid behavioral response. If the initial abnormal feature is that the body sinks vertically, and the posture remains unchanged after the tactile stimulus, then it is judged as no valid behavioral response. The response analysis unit (33) outputs a three-state judgment result: valid response, invalid response, and uncertain (when behavioral data cannot be fully collected due to occlusion or noise).

[0043] The risk update unit (34) receives the judgment result from the response analysis unit (33) and updates the risk level. When the judgment result is a valid response, the risk level is reduced or the warning is lifted, and the monitored person is continuously monitored during the preset attention period after the warning is lifted. When the judgment result is an invalid response, the risk level is maintained or upgraded. When the judgment result is uncertain, the current level is maintained, and the challenge decision unit (32) is triggered to issue a tactile stimulus signal again. The risk update unit (34) also feeds back the response judgment result to the initial detection unit (31) to form a closed loop. Specifically, based on the historical challenge-response record of the monitored person, the response criterion threshold and the initial detection threshold of the monitored person are adaptively adjusted. For example, for monitored persons who are frequently triggered with the initial risk signal in the same scenario but can quickly generate a valid response each time (such as professional swimmers who often do long-term breath-holding training), the system gradually adjusts the initial detection threshold of the individual—for example, extending the alarm duration threshold for the head submerged in the water from 15 seconds to 20 seconds—to reduce unnecessary challenge triggers. Conversely, for monitored individuals whose historical records show no response to the first level challenge but a response only to the second level challenge, the system lowers the response criterion threshold for that individual, making the judgment of behavioral changes more sensitive and avoiding misjudgment as no response due to the individual's insensitivity to vibration. This adaptive optimization based on individual historical data enables the system to achieve personalized detection sensitivity for different monitored individuals after long-term operation.

[0044] The alarm module (40) executes tiered alarms based on the updated risk level. Level 1 warnings are triggered when the initial risk signal is generated, sending a vibration alert and the location information of the monitored person to the receiving device worn by the lifeguard, without triggering a public alarm. Level 2 warnings are triggered when the risk escalates after verification, triggering an audible and visual alarm, causing strong vibrations in the lifeguard's equipment, and displaying the location of the monitored person and real-time footage in a pop-up window on the monitoring screen. Level 3 warnings are the highest level, linking the emergency broadcast system, locking the location of the nearest lifeguard, and automatically starting video recording and archiving. All AI inference is completed locally in the processing module without uploading to the cloud, achieving low-latency processing.

[0045] This invention provides a drowning detection verification method based on human-factor closed-loop feedback, the process of which is as follows: Figure 2 As shown, the closed-loop verification timing is as follows: Figure 3 As shown, it includes the following steps.

[0046] Step S1, Initial Detection. The sensor module (10) continuously collects behavioral data of the monitored person. The initial detection unit (31) of the processing module (30) analyzes the behavioral data. When a feature that meets the preset abnormal conditions is detected, an initial risk signal is generated. Figure 3At time T0, the video stream captured by the surface camera is processed by the YOLO target detection model to identify the position and posture of each monitored person. The DeepSORT algorithm performs multi-target tracking on each monitored person to generate a motion trajectory. An initial risk signal is generated when any of the following conditions are met: the head remains submerged for more than a preset time (e.g., 15 seconds for adults, 10 seconds for children); the body is in a vertical sinking posture for more than a preset time; the limb movement frequency matches a preset drowning struggle pattern (e.g., frequency in the range of 0.5-2Hz and irregular); the motion trajectory suddenly changes from a moving state to a stationary state for more than a preset time; or underwater sonar detects abnormal acoustic signatures matching drowning behavior. The initial risk signal includes the monitored person's identification, risk level (low / medium / high), and a description of abnormal features. Simultaneously, a level one warning is triggered when the initial risk signal is generated, sending a vibration alert and the monitored person's location information to the lifeguard's wristband. The lifeguard begins to pay attention to the monitored person but does not trigger a public alarm.

[0047] Step S2, tactile challenge. The challenge decision unit (32) receives the initial risk signal and determines the challenge strategy. For low / medium risk, the first-level challenge is adopted, with short, mild vibration (lasting 0.5 seconds, intensity of 50% of the motor's rated power); for high risk, the first-level challenge is adopted, with continuous strong vibration (lasting 1 second, intensity of 100% of the motor's rated power). The challenge instruction is sent wirelessly to the wearable device (20) of the monitored person. After receiving the instruction, the wearable device (20) drives the vibration motor to emit tactile stimulation signals, corresponding to... Figure 3 At time T1 in the middle.

[0048] Step S3, Response Observation. Within a preset time window (preferably 3-5 seconds) after the tactile stimulus signal is emitted, the sensor module (10) continuously collects the behavioral data of the monitored person as response observation data, corresponding to... Figure 3 During the observation time window from T1 to T1+Δt, the camera continuously collects video frames, the sonar continuously collects acoustic data, and the wristband motion sensor continuously collects acceleration and angular velocity data.

[0049] Step S4, Behavioral Response Analysis. The response analysis unit (33) compares the response observation data with the behavioral data before the tactile stimulus signal is emitted to determine whether the monitored person has produced an effective behavioral response. Specifically, the determination is made using one or a combination of the three methods mentioned above. For example, when using method one (time-series comparison), the Euclidean distance between the behavioral feature vectors 3 seconds before and 3 seconds after the challenge is calculated. If the distance exceeds a preset threshold (such as 3 times the standard deviation of the fluctuation of the behavioral feature vector when the monitored person is swimming normally), it is determined to be an effective behavioral response. The behavioral feature vector includes multimodal data from the camera, sonar, and wristband motion sensor. When using method two (specific action recognition), the video frame after the challenge is input into the pre-trained action classification model. If the probability of the model outputting responsive actions such as raising the head to breathe or changing direction exceeds 0.8, it is determined to be an effective behavioral response. At the same time, if the acceleration and angular velocity data of the wristband motion sensor show pattern changes that match responsive actions such as raising the wrist or turning, it is also used as an auxiliary criterion for effective behavioral response. When using method three (signal persistence analysis), check whether the initial abnormal feature has disappeared. For example, if the initial abnormal feature is that the head is submerged in the water, and after the challenge, the camera detects that the head is floating out of the water and the data from the wristband motion sensor changes from a near-zero state to a continuous fluctuation that matches the swimming motion, then it is determined to be a valid behavioral response.

[0050] Step S5, Risk Status Update and Alarm Output. Update the risk level based on the behavioral response judgment result, corresponding to... Figure 3 At time T2, if the response is deemed valid, the risk level is lowered or the alert is lifted, the event log is recorded (including initial anomaly characteristics, challenge parameters, and response characteristics), and continuous monitoring begins. If the response is deemed invalid, the process proceeds as follows: Figure 4 The multi-level challenge process shown in the diagram issues a second-level tactile stimulus signal (strong vibration, lasting 2 seconds), and executes steps S3 and S4 again for a second response judgment. If the second response is still invalid, it is upgraded to a level two warning, triggering an audible and visual alarm. If the second response is valid, it is downgraded to a level one warning for continued observation. If there are invalid responses after two consecutive levels of challenges and the initial risk is high, it is upgraded to a level three warning. When the judgment is uncertain, the current risk level is maintained, and a second challenge is triggered (the tactile stimulus signal is issued again and steps S3-S4 are repeated). The response judgment result is simultaneously fed back to the initial detection unit (31) to adjust subsequent detection parameters (such as fine-tuning the abnormal condition threshold of the monitored person), forming a closed loop. The closed loop feedback includes: adaptively adjusting the response criterion threshold and the initial detection threshold of the monitored person based on the historical challenge-response record of the monitored person, so as to achieve personalized detection sensitivity optimization.

[0051] In one application scenario, taking a standard 50m×25m swimming pool as an example, the working process of this invention in two typical scenarios is illustrated. The pool is equipped with four 4K surface cameras and eight underwater sonar sensors. Those being monitored receive a waterproof wristband (containing a built-in three-axis accelerometer, three-axis gyroscope, water immersion sensor, vibration motor, and LoRa communication module) upon entry and return it upon exit. Two lifeguards wear lifeguard wristbands.

[0052] For example, the behavior of the monitored person meets the preset abnormal conditions but there is no actual risk of drowning. A monitored person is conducting breath-holding diving training in a deep water area, with their head submerged in the water for more than 15 seconds and their body almost still. The camera detects the feature of the head being continuously submerged in the water and the body being still, generating a level one risk signal (low risk). The lifeguard wristband receives a vibration reminder. The system sends a level one tactile stimulus (short, mild vibration) through the monitored person's wristband. After feeling the vibration, the monitored person raises their head to breathe and swims towards the shore. The camera captures the response features of the head surfacing and the movement trajectory changing from stillness to horizontal movement within a 3-second observation window. At the same time, the wristband motion sensor detects that the acceleration and angular velocity data change from near zero to continuous fluctuations matching the swimming movements. The response analysis unit (33) uses method three combined with method one to judge: the initial abnormal feature (head submerged in the water) has disappeared, and the change amplitude of the behavior feature vector exceeds the preset threshold, which is judged as a valid behavior response. The system cancels the warning, records the event log, and the lifeguard wristband stops vibrating. In this scenario, false alarms were eliminated, no public alarms were issued, the experience of the monitored person was not affected, and the lifeguard's attention was not unnecessarily diverted.

[0053] For example, a person being monitored sinks in deep water due to leg cramps. The camera detects characteristics such as vertical sinking and irregular limb struggle, and the sonar detects abnormal underwater acoustic patterns, generating a level 2 risk signal (high risk). The lifeguard's wristband receives a vibration alert. The system sends a level 1 tactile stimulus (strong vibration) through the monitored person's wristband. The monitored person has lost the ability to act independently due to drowning and is unable to make a conscious behavioral response. Within a 3-second observation window, the camera does not detect any behavioral changes (the person remains in a sinking posture), the abnormal acoustic patterns of the sonar continue, and the wristband motion sensor data does not show any acceleration or angular velocity change patterns corresponding to the monitored person's active movement. The response analysis unit (33) judges that: the behavioral feature vector has no significant change (method 1), the initial abnormal features continue to exist (method 3), and it is determined to be an invalid behavioral response. The system emits a second-level tactile stimulus (stronger vibration, lasting 2 seconds). If no significant changes are observed in the sensor data within the 3-second observation window, the system upgrades to a third-level warning, triggering an audible and visual alarm. The lifeguard's wristband vibrates strongly and displays the monitored person's location coordinates, while a pop-up window on the monitoring screen displays a real-time image of the monitored person. In this scenario, the entire process from initial detection to the third-level alarm takes approximately 8-10 seconds. The lifeguard immediately locates and provides assistance, and the first-level warning was issued during the initial detection (second 0), indicating that the lifeguard has already begun monitoring the area.

[0054] In another application scenario, the challenge-response verification mechanism of this invention is not limited to drowning detection, but can be extended to other safety monitoring scenarios that require distinguishing whether abnormal signals are genuine. For example, in the field of fall detection for the elderly, when a wearable device detects a movement pattern that matches the characteristics of a fall, it generates an initial risk signal, and the system issues a vibration challenge through the device. If the elderly person has not fallen, they will generate a natural body movement response after feeling the vibration, which is determined to be a valid behavioral response and the warning is lifted; if the elderly person has indeed fallen and is unable to move, no valid body movement response will be generated, which is determined to be an invalid behavioral response and an emergency call will be triggered.

[0055] Furthermore, the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the above-described method steps S1 to S5. The computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk. The computer program may be deployed on the processor of a processing module for execution, or it may be deployed on a cloud server for execution (when the system architecture adopts a cloud-edge collaborative mode).

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A drowning detection verification method based on human-caused closed-loop feedback, characterized in that, Includes the following steps: S1. Detect the behavior data of the monitored person through one or more sensors, and generate an initial risk signal when the behavior data contains features that meet preset abnormal conditions. S2. In response to the initial risk signal, a tactile stimulation signal is emitted to the monitored person through the wearable device worn by the monitored person; S3. Within a preset time window after the tactile stimulation signal is emitted, the behavioral data of the monitored person is continuously collected by the sensor as response observation data. S4. The system compares the response observation data with the behavioral data before the tactile stimulus signal is emitted, and automatically determines whether the monitored person produces an effective behavioral response without relying on the monitored person to perform manual operation; wherein, the effective behavioral response refers to the behavioral changes produced by the monitored person within the preset time window that indicate that they have consciousness and action capabilities. S5. Update the risk level based on the judgment result of the effective behavior response, and output the alarm information corresponding to the updated risk level.

2. The method according to claim 1, characterized in that, The criteria for an effective behavioral response include at least one of the following: The head position of the monitored person changed from being submerged in the water to being floating on the surface; The monitored individual's movement characteristics changed from an abnormal pattern to a normal swimming pattern; The acceleration or angular velocity data of the wearable device exhibits characteristic changes that match preset responsive actions.

3. The method according to claim 1, characterized in that, Step S4 determines whether the monitored subject produces a valid behavioral response, including at least one of the following methods: The behavioral feature vector within the first time window before the tactile stimulus signal is emitted is compared with the behavioral feature vector within the second time window after the emission. When the change amplitude of the behavioral feature vector exceeds a preset threshold, it is determined to be a valid behavioral response. The system uses a pre-trained machine learning model to identify whether the monitored person exhibits a preset responsive action after the tactile stimulus signal is emitted. The responsive action includes at least one of raising the head to breathe, changing swimming direction, and raising the wrist to check the device. The abnormal behavioral characteristics corresponding to the initial risk signal are analyzed to see whether they disappear or weaken after the tactile stimulus signal is issued. If they disappear or weaken, they are determined to be a valid behavioral response.

4. The method according to claim 1, characterized in that, The tactile stimulation signal includes at least one level of tactile stimulation, and the method further includes a multi-level challenge step: If no effective behavioral response is detected within the preset time window after the first-level tactile stimulation signal is emitted, a second-level tactile stimulation signal is emitted and a second behavioral response judgment is made, wherein the intensity of the first-level tactile stimulation signal is lower than the intensity of the second-level tactile stimulation signal. If no effective behavioral response is detected after multiple levels of tactile stimulation, the risk level is upgraded to the highest level.

5. The method according to claim 1, characterized in that, The one or more sensors include at least one of the following: Image acquisition equipment deployed above the water surface is used to collect visual posture data and motion trajectory data of the monitored subject; Acoustic sensors deployed underwater are used to collect underwater acoustic data; The wearable device has a built-in motion sensor for collecting acceleration and angular velocity data of the monitored person.

6. The method according to claim 1, characterized in that, Step S5 involves updating the risk level, including: When a valid behavioral response is detected, the risk level is reduced or the warning is lifted, and the monitored person is continuously monitored for a preset monitoring period after the warning is lifted. If no effective behavioral response is detected, maintain or upgrade the risk level. When behavioral data cannot be fully collected due to occlusion or noise, resulting in an uncertain response judgment, the current risk level is maintained and a tactile stimulus signal is triggered again.

7. The method according to claim 1, characterized in that, It also includes a closed-loop feedback step: The judgment result of the effective behavioral response is fed back to the initial detection step to dynamically adjust the threshold parameters, risk model weights, or response criterion thresholds of the monitored individuals in subsequent detections. The response criterion thresholds of the monitored individuals are adaptively adjusted based on their historical challenge-response records.

8. The method according to claim 1, characterized in that, The preset time window is 2 to 10 seconds.

9. The method according to claim 1, characterized in that, The tactile stimulation signal is also accompanied by at least one of the following auxiliary stimulation signals: The light signal emitted by the wearable device; The wearable device emits sound signals.

10. A drowning detection and verification system based on human-cause closed-loop feedback, characterized in that, include: A sensor module, comprising one or more sensors, is used to detect behavioral data of the monitored person. The wearable device, worn by the monitored person, includes a tactile stimulation module and a wireless communication module, used to receive instructions from the processing module and send tactile stimulation signals to the monitored person; The processing module, communicatively connected to the sensor module and the wearable device, is used to generate an initial risk signal when the behavioral data contains features that meet preset abnormal conditions; in response to the initial risk signal, it sends a tactile stimulation signal to the monitored person through the wearable device; within a preset time window after sending the tactile stimulation signal, it continuously collects the monitored person's behavioral data as response observation data through the sensor module; and compares the response observation data with the behavioral data before the tactile stimulation signal was sent to determine whether the monitored person has produced an effective behavioral response. And update the risk level based on the judgment result of the effective behavioral response; The alarm module is communicatively connected to the processing module and is used to output corresponding alarm information according to the updated risk level. The determination of the effective behavioral response is automatically executed by the processing module, without relying on the monitored subject to perform manual operations.

11. The system according to claim 10, characterized in that, The wearable device also includes motion sensors, including accelerometers and gyroscopes, for collecting motion data of the monitored person and transmitting it to the processing module as part of the response observation data.

12. The system according to claim 10, characterized in that, The sensor module includes an image acquisition device deployed above the water surface and / or an acoustic sensor deployed underwater; the processing module also includes a multi-sensor fusion unit for time synchronization and spatial correlation of data from different sensors to achieve cross-validation.

13. The system according to claim 10, characterized in that, The processing module also includes a challenge decision unit, which is used to determine the level, vibration mode, duration and observation window length of the tactile stimulation signal based on at least one of the following: the level of the initial risk signal, the information of the monitored person and environmental conditions.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 9.

Images