Anti-drowning monitoring method and system based on AI

Abstract

The invention discloses an AI-based drowning prevention monitoring method and system. The system comprises a monitoring terminal, a central server, a video monitoring center and a mobile terminal. The monitoring terminal is composed of a high-definition network camera, an AI image edge calculation module, a microprocessor, a data transmission module and an alarm module, the microprocessor is connected with the AI image edge calculation module, the data transmission module and the alarm module, and the AI image edge calculation module is connected with the high-definition network camera. According to the system, millisecond-level analysis is realized through an edge computing architecture, infrared thermal imaging and underwater sensing technologies can be combined, day and night, weather and water quality limitations can be broken through, a golden rescue window is prolonged, multi-source data cross validation is realized through fusion of multi-modal data based on skeleton key point tracking and motion entropy analysis, and short-time-sequence window analysis is performed through an edge layer. Local sound-light alarm is carried out to strive for life-saving time, and the central server carries out long time sequence analysis so as to reduce the probability of misjudgment and improve the accuracy.

Description

Technical Field

[0001] The present invention relates to the field of AI intelligent security monitoring technology, and in particular to an AI-based drowning prevention monitoring method and system. Background Art

[0002] In the high temperature environment in summer, it is common for children to swim in lakes, reservoirs and rivers without permission, especially teenagers and children living near lakes, reservoirs and rivers, resulting in many drowning accidents.

[0003] Even in places like swimming pools and water parks that are equipped with safety officers, the officers themselves have limited energy and there are too many people to always pay attention to every swimmer. Therefore, drowning accidents occasionally occur in swimming pools and water parks.

[0004] Therefore, in order to solve the above problems, the present invention proposes a device for implementing real-time safety monitoring of lakes, rivers, swimming pools and other places based on AI technology. First, a large model is generated by AI, and various normal swimming methods and states of human swimming are used as training models to generate a variety of swimming data models such as freestyle, backstroke, breaststroke, butterfly stroke, and diving, which are mainly used to distinguish between drowning states and natural swimming. Then, the swimming characteristics of swimmers within the monitoring range are captured in real time by a camera device, and analyzed online in real time to determine whether the swimmer is in a safe swimming state or in an unsafe state such as drowning. If a swimmer is found to be in a dangerous state, the device will draw a drowning conclusion from the operation center, send the information to the controller and the information transceiver, start the alarm system, make an on-site alarm and transmit the alarm information to the control center, waiting for the on-site personnel or the control center to confirm whether the swimmer is safe, so as to eliminate the safety hazards of the swimmer. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an AI-based anti-drowning monitoring system, thereby realizing AI intelligent drowning monitoring of water areas, so that drowning people can be discovered in time, and relevant personnel can be quickly notified for rescue.

[0006] The technical solution of the present invention is:

[0007] An AI-based drowning prevention monitoring system includes a monitoring terminal, a central server, a video monitoring center, and a mobile terminal. The monitoring terminal is composed of a high-definition network camera, an AI image edge computing module, a microprocessor, a data transmission module, and an alarm module. The microprocessor is connected to the AI ​​image edge computing module, the data transmission module, and the alarm module. The AI ​​image edge computing module is connected to the high-definition network camera.

[0008] There are multiple monitoring terminals, which are installed off the ground in the center or at the edge of the water surface and connected to a central server. The central server is connected to a video monitoring center and is wirelessly connected to a number of mobile terminals.

[0009] Furthermore, the high-definition network camera of the monitoring terminal transmits the image to the AI ​​image edge computing module for primary image processing and judgment. The microprocessor activates the alarm module based on the judgment result and transmits the image data to the central server through the data transmission module for secondary processing and judgment.

[0010] Furthermore, the monitoring terminal is also connected to auxiliary cameras, which include underwater cameras and infrared / thermal imaging cameras.

[0011] Furthermore, the AI ​​image edge computing module adopts the NVIDIA Jetson series or Intel NUC series.

[0012] The present invention also provides a drowning prevention monitoring method based on the above monitoring system, comprising the following steps:

[0013] (1) Target detection and tracking: Detect human targets in each frame, assign unique IDs to detected targets, track them across frames, establish the "lifecycle" of each target, and build a swimming model database through AI machine learning and deep learning;

[0014] (2) Single-frame feature extraction: for each tracked target, skeleton key points are detected and key drowning indicators are calculated in each frame, and the above data are analyzed and compared with the AI ​​large model database;

[0015] (3) Short time window analysis: within N consecutive frames, the indicators of each target are statistically analyzed using the AI ​​big model and enter the analysis phase;

[0016] (4) Drowning probability scoring: Based on the results of statistical analysis of the AI ​​large model, a weighted scoring model is designed to assign weights to each feature according to its importance;

[0017] (5) One-time decision and alarm: if the risk score of a target in the current short window exceeds the preset higher threshold T1, it is considered that the possibility of drowning is extremely high, and the local sound and light alarm is immediately triggered.

[0018] (6) Secondary decision-making and alarm: observe the target's behavior pattern over a longer period of time. If the risk score remains high or exceeds a slightly lower threshold T2, confirm the edge alarm, trigger the management-level alarm, and notify the mobile terminal. If the long-term analysis shows that the target behavior has returned to normal, overrule the edge alarm.

[0019] Furthermore, key drowning indicators include vertical body position, head posture, arm movement, leg movement, movement trajectory and speed, and immersion time.

[0020] Furthermore, each feature is assigned a weight based on its importance, specifically including:

[0021] Vertical position, the observation index is the angle θ between the shoulder-hip line and the horizontal plane, which lasts for more than 5 seconds and has a weight of 0.2;

[0022] Head posture, the observation indicator is the proportion of time the mouth and nose are underwater, lasting for a 5-second window, with a weight of 0.25;

[0023] For arm movement, the observation index is the disorder of arm movement, with a weight of 0.2;

[0024] Motion trajectory and speed: the observation indicators are the proportion of frames with speed < 0.1 m / s + trajectory chaos, with a weight of 0.15;

[0025] Leg movement: the observation indicator is the proportion of frames with leg movement amplitude <15px, 5-second window, and weight 0.1;

[0026] Immersion time: The observation indicator is the time it takes for the target to disappear from the water surface, with a weight of 0.1.

[0027] The present invention is beneficial in that:

[0028] (4) The system achieves millisecond-level analysis through edge computing architecture and triggers local sound and light alarms within 5 seconds, which is three times faster than the response speed of traditional manual monitoring. It can combine infrared thermal imaging and underwater sensing technology to break through the limitations of day and night, weather, and water quality, and extend the golden rescue window.

[0029] (5) Based on skeletal key point tracking and motion entropy analysis, by fusing multimodal data and realizing multi-source data cross-validation, the system has a recognition rate of 92% for 9 types of drowning postures and a false alarm rate as low as 21%, solving the problem of poor environmental adaptability of traditional solutions.

[0030] (6) The system builds an intelligent hierarchical response system. The edge layer performs short-term window analysis and issues local sound and light alarms to save time. The central server performs long-term analysis (30-60s) to observe the target's behavior pattern over a longer period of time, reducing the probability of misjudgment and improving accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the system architecture of the present invention;

[0032] Figure 2 This is a schematic diagram of the monitoring terminal system architecture;

[0033] Figure 3Flowchart of the monitoring method in the present invention. DETAILED DESCRIPTION

[0034] The following is a further description of specific embodiments of the present invention in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is intended to facilitate understanding of the present invention and does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0035] As shown in the figure:

[0036] An AI-based drowning prevention monitoring system includes a monitoring terminal, a central server, a video monitoring center, and a mobile terminal. The monitoring terminal is composed of a high-definition network camera, an AI image edge computing module, a microprocessor, a data transmission module, and an alarm module. The microprocessor is connected to the AI ​​image edge computing module, the data transmission module, and the alarm module. The AI ​​image edge computing module is connected to the high-definition network camera.

[0037] There can be multiple monitoring terminals, which are installed off the ground in the center or edge of the water surface and connected to a central server. The central server is connected to a video surveillance center and wirelessly connected to several mobile terminals. The specific selection can be based on the conditions of the water area. For example, 1-2 can be installed in a swimming pool. In order to increase the accuracy of monitoring, the monitoring terminal can also be connected to auxiliary cameras. The auxiliary cameras include underwater cameras and infrared / thermal imaging cameras. The underwater cameras can provide underwater perspectives and reduce interference from surface reflections. They are especially effective for drowning people. However, the deployment and maintenance costs are high and they can be used in small-scale waters. The infrared / thermal imaging cameras can detect human body heat characteristics in dark, smoky or turbid waters and can be used in larger natural waters.

[0038] In the present invention, the high-definition network camera of the monitoring terminal transmits the image to the AI ​​image edge computing module (NVIDIA Jetson series or Intel NUC series is recommended) for primary image processing and judgment. The microprocessor activates the alarm module according to the judgment result and transmits the image data to the central server through the data transmission module for secondary processing and judgment.

[0039] The AI ​​image edge computing module has the following functions:

[0040] (1) Receive and process camera video streams in real time, and run lightweight object detection models to detect and locate all human targets in the picture in real time (YOLOv5 / v8 Nano / Tiny, EfficientDet-Lite, MobileNet SSD).

[0041] (2) Run the key point detection model to detect the skeleton key points of the human body (such as OpenPose Lite, MoveNet, MediaPipe BlazePose), which is the core of posture judgment.

[0042] (3) Preliminary drowning feature extraction, calculation of real-time indicators based on key points (see the method section below), and making a decision and alarm.

[0043] Local real-time alarm triggering: When the preliminary algorithm determines that the possibility of drowning exceeds a high threshold, it immediately triggers a local sound and light alarm (such as a high-decibel alarm by the swimming pool, a strong strobe light, etc.) to gain golden rescue time. This is the core value of edge computing - low latency, which not only significantly reduces network bandwidth requirements, reduces transmission delays, and improves system response speed, but also allows local alarms to continue working even if the network is interrupted.

[0044] The central server’s functions are as follows:

[0045] (1) Receive and store data uploaded by the AI ​​image edge computing module, run more complex drowning analysis algorithms, use larger and more accurate models (such as Transformer-based models) for multi-target tracking, long-term behavior analysis, and multi-feature fusion judgment, integrate information from multiple camera perspectives, and reduce occlusion misjudgments.

[0046] (2) Secondary confirmation and reduction of false alarms: By combining behavioral patterns over a longer time window (e.g., 30-60 seconds) and data from multiple monitoring terminals, the primary alarm at the edge layer can be confirmed or rejected, thereby improving the accuracy of the early warning.

[0047] (3) Alarm management: After drowning is confirmed, a higher-level alarm is triggered: notification to mobile terminals, such as the pool administrator app / computer pop-up window, sending text messages / calls to lifeguards, calling pagers, mobile phones, etc., linking the broadcast system, and recording video clips of the incident.

[0048] System management: user management, device management (camera status monitoring), alarm log query, data analysis (drowning hotspots, time analysis), model update deployment, etc.

[0049] The specific monitoring methods are as follows:

[0050] Step 1: Target Detection and Tracking

[0051] Human targets are detected in each frame of the image, and the detected targets are assigned unique IDs and tracked across frames (using algorithms such as SORT and DeepSORT). The "lifecycle" of each target is established, and a swimming model database is constructed through AI machine learning and deep learning.

[0052] Step 2: Single-frame features

[0053] For each tracked target, skeleton key points are detected and key drowning indicators are calculated in each frame. The above data is analyzed and compared with the AI ​​large model database. Key drowning indicators include:

[0054] a. Vertical position

[0055] Principle: Drowning people are usually unable to perform effective treading water, and their bodies tend to float vertically in the water, which is in sharp contrast to the swimmer's usual horizontal posture (head forward). This is one of the core characteristics.

[0056] Algorithm: Calculate the angle between the line connecting key points of the human body (such as shoulders and hips) and the horizontal plane. A sustained angle (such as >5 seconds) close to 90 degrees is a strong indicator.

[0057] b. Head posture

[0058] Principle: Even when a drowning person is struggling, his mouth and nose are frequently or continuously submerged in water.

[0059] Algorithm: Detect the position of nose and mouth key points relative to the waterline (may require waterline detection or fixed area assumption). Calculate the head angle (chin pointing).

[0060] c. Arm exercise

[0061] Principle: Drowning people usually slap the water surface with their arms at the sides or in front of the body irregularly and quickly, without effective paddling propulsion and periodic effective movements. At the same time, drowning people sometimes try to push the water down with their arms horizontally or stretched downward to the sides of the body.

[0062] Algorithm: Tracks the motion trajectory of key points of the hand, analyzes the amplitude, frequency, direction (large vertical component), regularity (weak or chaotic periodicity), and calculates the angle between the arm and the torso.

[0063] d. Leg exercises

[0064] Principle: The legs of drowning people usually lack effective treading or kicking movements and appear stiff or weak.

[0065] Algorithm: Track the movement trajectory of key points of the foot and analyze the movement amplitude and frequency.

[0066] e. Movement trajectory and speed

[0067] Principle: Drowning people are usually unable to move effectively, struggling in place in the water or drifting slowly with the current.

[0068] Algorithm: Perform multi-target tracking on detected human targets, and calculate their movement speed (very low) and the degree of chaos of their movement trajectory.

[0069] f. Immersion time

[0070] Principle: The drowning person is submerged for a long time, sinking underwater for more than a certain time (such as >20 seconds, the specific threshold is adjustable) without obvious surfacing or effective movement.

[0071] Algorithm: Start timing when the target disappears from the water surface (cannot be detected or the key point is below the waterline).

[0072] Step 3: Short time window analysis

[0073] Within N consecutive frames (e.g., 5-10 seconds), the AI ​​big model is used to perform statistical analysis on the indicators of each target and enter the analysis phase;

[0074] Step 4: Drowning Probability Score

[0075] Based on the analysis results of the AI ​​big model, a weighted scoring model is designed to assign weights to each feature according to its importance. The following is a designed weighted scoring model.

[0076]

[0077] Note:

[0078] Motion entropy calculation: Shannon entropy (8-directional partitioning) is calculated through the histogram of the movement directions of the key points of the hand. The higher the entropy value, the more disordered the movement.

[0079] Trajectory confusion: Calculate the variance of the target center point's moving direction (0°-360°). A variance > 90° will result in a score of 100.

[0080] Special scene weight adjustment:

[0081]

[0082]

[0083] The present invention can generate a large model through AI, and use various normal swimming methods and states of human swimming as training models for the above scoring model, to generate various swimming data models such as freestyle, backstroke, breaststroke, butterfly stroke, and diving, which are mainly used to distinguish between drowning states and natural swimming. The model is then continuously trained to record more swimming posture data, so as to provide more accurate early warning and judgment.

[0084] Step 5: Decision-making and alarm

[0085] If the risk score of a target within the current short window (5-10 seconds) exceeds the preset higher threshold T1 (85 points), the possibility of drowning is considered extremely high, and a local sound and light alarm is immediately triggered. This step is performed by the edge decision layer composed of monitoring terminals.

[0086] Step 6: Secondary decision and alarm

[0087] Observe the target's behavior pattern over a longer period of time (30-60 seconds). If the risk score remains high or exceeds the slightly lower threshold T2 (65 points), confirm the edge alarm, trigger the management-level alarm, and notify the mobile terminal. If the long-term analysis shows that the target behavior has returned to normal, overrule the edge alarm. This process is performed by the central server, with the purpose of reducing misjudgments and improving accuracy.

[0088] This solution adopts a layered strategy to balance speed and accuracy. The edge layer focuses on high-confidence features (vertical body position + nasal and mouth immersion), while the central server reduces false positives through long-term time series analysis. Actual deployment requires collecting local pool data to continuously optimize thresholds. Initially, it is recommended to run the system for 72 hours under lifeguard supervision to calibrate parameters.

[0089] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It is apparent to those skilled in the art that various changes, modifications, substitutions, and variations to these embodiments may be made without departing from the principles and spirit of the present invention, and these changes and modifications still fall within the scope of protection of the present invention.

Claims

1. An AI-based drowning prevention monitoring system, characterized by: It includes a monitoring terminal, a central server, a video monitoring center and a mobile terminal; the monitoring terminal is composed of a high-definition network camera, an AI image edge computing module, a microprocessor, a data transmission module and an alarm module. The microprocessor is connected to the AI ​​image edge computing module, the data transmission module and the alarm module, and the AI ​​image edge computing module is connected to the high-definition network camera; There are multiple monitoring terminals, which are installed off the ground in the center or at the edge of the water surface and connected to a central server. The central server is connected to a video monitoring center and is wirelessly connected to a number of mobile terminals.

2. The AI-based drowning prevention monitoring system according to claim 1, characterized in that: The high-definition network camera of the monitoring terminal transmits the image to the AI ​​image edge computing module for primary image processing and judgment. The microprocessor activates the alarm module according to the judgment result and transmits the image data to the central server through the data transmission module for secondary processing and judgment.

3. The AI-based drowning prevention monitoring system according to claim 1, characterized in that: The monitoring terminal is also connected to auxiliary cameras, which include underwater cameras and infrared / thermal imaging cameras.

4. The AI-based drowning prevention monitoring system according to claim 1, characterized in that: The AI ​​image edge computing module adopts the NVIDIA Jetson series or Intel NUC series.

5. A drowning prevention monitoring method based on the AI ​​drowning prevention monitoring system according to any one of claims 1 to 4, characterized in that: The following steps are involved: (1) Target detection and tracking: Detect human targets in each frame, assign unique IDs to detected targets, track them across frames, establish the "lifecycle" of each target, and build a swimming model database through AI machine learning and deep learning; (2) Single-frame feature extraction: for each tracked target, skeleton key points are detected and key drowning indicators are calculated in each frame, and the above data are analyzed and compared with the AI ​​large model database; (3) Short time window analysis: within N consecutive frames, the AI ​​big model is used to perform statistical analysis on the indicators of each target; (4) Drowning probability scoring: Based on the results of statistical analysis of the AI ​​large model, a weighted scoring model is designed to assign weights to each feature according to its importance; (5) One-time decision and alarm: if the risk score of a target in the current short window exceeds the preset higher threshold T1, it is considered that the possibility of drowning is extremely high, and the local sound and light alarm is immediately triggered. (6) Secondary decision-making and alarming: observe the target’s behavior pattern over a longer period of time. If the risk score remains high or exceeds a slightly lower threshold T2, a marginal alarm is confirmed, a management-level alarm is triggered, and the mobile terminal is notified. If the long-term analysis shows that the target behavior has returned to normal, overrule the edge alarm.

6. The drowning prevention monitoring method according to claim 5, characterized in that: The key drowning indicators include vertical body position, head posture, arm movement, leg movement, movement trajectory and speed, and immersion time.

7. The drowning prevention monitoring method according to claim 6, characterized in that: Assigning weights to each feature based on its importance specifically includes: Vertical position, the observation index is the angle θ between the shoulder-hip line and the horizontal plane, which lasts for more than 5 seconds and has a weight of 0.2; Head posture, the observation indicator is the proportion of time the mouth and nose are underwater, lasting for a 5-second window, with a weight of 0.25; For arm movement, the observation index is the disorder of arm movement, with a weight of 0.2; Motion trajectory and speed: the observation indicators are the proportion of frames with speed < 0.1 m / s + trajectory chaos, with a weight of 0.15; Leg movement: the observation indicator is the proportion of frames with leg movement amplitude <15px, 5-second window, and weight 0.1; Immersion time: The observation indicator is the time it takes for the target to disappear from the water surface, with a weight of 0.1.

Images