Treadmill fall early warning and emergency stop method and system based on posture recognition

Abstract

The present application relates to treadmill safety control technology, and proposes a kind of method and system for tumble early warning and emergency stop of treadmill based on posture recognition.The camera is arranged above the front of the treadmill, real-time user motion video is collected, human posture recognition algorithm is used to extract skeletal key points, body center of gravity position, body center of gravity speed, torso inclination and other posture features are calculated, and running parameters such as running belt speed, acceleration and safety area length are fused to construct imbalance risk index and predict remaining time.The system jointly determines the imbalance risk index and the predicted remaining time based on the adaptive threshold, triggers the sound-light or voice warning in the near-falling state and automatically slows down according to the preset curve, and when detecting the characteristics such as sudden drop of body center of gravity height, abnormal inclination of torso and serious loss of key points and meeting the emergency determination condition, the emergency stop command is issued to the drive controller to realize rapid braking, and relevant data is recorded for model updating.

Description

Technical Field

[0001] This invention relates to treadmill safety control technology, and more particularly to a method and system for treadmill fall warning and emergency stop based on posture recognition. Background Technology

[0002] With the widespread use of treadmills in home fitness, commercial gyms, and rehabilitation training settings, safety issues during treadmill use are becoming increasingly prominent. Treadmills are typical active devices where high-speed moving parts directly interact with the human body. If a user missteps, trips, or loses balance and is carried out of the safe zone by the conveyor belt, they are prone to secondary falls, abrasions, or even serious injuries from collisions with other hard objects. Therefore, how to provide timely warnings when a user's posture is abnormal or they are about to fall, and how to quickly and controllably stop the machine when necessary, has become an important research direction for treadmill safety control technology.

[0003] Most treadmills currently available are equipped with mechanical safety ropes or magnetic safety keys. These keys are attached to the user's clothing via a thin wire. When the user falls or leaves the machine, pulling the rope breaks the key, detaching it from the sensor and cutting off the control circuit for an emergency stop. While this device is simple and inexpensive, it has significant limitations: First, it relies on the user actively attaching the safety rope, which many users don't wear due to inconvenience, arm movement restrictions, or other reasons, rendering the safety device ineffective. Second, it only passively cuts off the power after a clear separation has occurred (usually after the person has been thrown off the running belt), failing to intervene in the pre-fall imbalance stage. Third, it cannot recognize the user's posture, relying solely on whether the rope has broken to determine the state, making it impossible to distinguish between a normal dismount and a dangerous fall, or to implement different deceleration or alarm strategies based on varying levels of risk. US Patent 10286286B1 discloses a treadmill safety device that incorporates ultrasonic or optical sensors to measure the distance between the runner and the treadmill, building upon a traditional magnetic safety key. When the device detects that the user has moved away from a predetermined safety area, it controls the treadmill to slow down or stop, addressing the issue of users being unwilling to use safety harnesses. However, this solution still relies solely on the "distance to the front of the conveyor belt," failing to recognize changes in the user's posture, torso tilt, and gait imbalance on the belt. It also struggles to accurately distinguish normal actions such as brief backward movements or turning around to communicate, potentially leading to false stops or missed detections. Furthermore, this solution focuses on distance detection and safety link improvements, lacking modeling of dynamic human posture characteristics and failing to establish a linkage with treadmill control parameters (speed, acceleration).

[0004] To reduce the risk of falls among high-risk groups such as the elderly, fall detection technology that integrates vision and sensors has developed rapidly in recent years. Chinese invention patent CN111383421A discloses a privacy-preserving fall detection method and system. This method uses a camera to capture human video, extracts key points of the human skeleton through a posture estimation module, and combines this with a behavior recognition module to classify postures into predefined behavioral sequences such as standing, sitting, and lying down. The fall detection module then analyzes these behavioral sequences based on a state machine to determine whether a fall has occurred. To protect privacy, the system performs posture estimation and behavior recognition on a local embedded device, sending only "de-featured" skeletal images and alarm information to the server, thus avoiding the uploading of raw video data. This type of technology can already achieve relatively reliable "fall detection-alarm" in scenarios such as indoor elderly care, while also considering a certain degree of privacy protection. However, its design goal is mainly for relatively static scenarios such as rooms and corridors, dividing the ground area into dangerous / normal zones. The core focus is on how to issue an alarm as quickly as possible after a fall occurs, rather than performing process risk assessment and linkage control on high-speed moving equipment. In addition, such systems typically do not require real-time interaction with the drive control system of specific motion equipment and lack deep coupling with motor control, motion curve adjustment, emergency stop logic, etc.

[0005] In the field of sports and fitness, several vision-based fall detection solutions have been proposed. Chinese invention patent CN114783051A discloses a vision-based method for fall detection and alarm in outdoor fitness running scenarios. This method uses a monocular camera deployed on outdoor public fitness trails to capture RGB images of runners. Human detection is performed on the images to obtain detection boxes, and then human posture estimation is performed to obtain 24 skeletal key points. These key points are converted into feature vectors using a motion posture coding model, constructing a fall feature library and a non-fall feature library. During online monitoring, the extracted posture features are compared with the feature libraries for similarity. An alarm is triggered when a preset threshold condition is met and the fall persists for a certain period, thus enabling fall detection on outdoor fitness trails. This technology incorporates skeletal key points and posture coding, improving the accuracy of runner fall detection in outdoor sports scenarios and making it suitable for safety monitoring of open trails. However, the solution still only issues an alarm after confirming that the person has "fallen to the ground," and it is not specifically designed for closed equipment scenarios such as treadmills. On the one hand, it does not participate in the real-time control of the treadmill drive system and cannot dynamically adjust the running belt speed or perform graded braking according to the degree of risk in the early stages of abnormal posture. On the other hand, outdoor running tracks do not have unique danger areas such as conveyor belt boundaries and equipment platforms, and the solution does not consider integrating information such as the relative position and speed of the running belt to predict the risk of being thrown off the running belt.

[0006] Generally speaking, the existing fall detection technologies can be roughly divided into three categories: sensor solutions centered around wearable devices such as accelerometers and gyroscopes, vision solutions centered around depth cameras or ordinary RGB cameras, and simple state detection solutions based on distance or pressure sensors. The sensor solutions require the monitored person to continuously wear the device, with poor compliance and easy to cause restraint in exercise scenarios such as high-intensity running; although the vision solutions can non-contact identify fall behaviors, most of them are for indoor elderly care or general scenarios, paying more attention to rapid alarm and privacy protection after a fall rather than linkage control on exercise equipment; the treadmill safety devices based on distance or contact status often only use a single physical quantity (such as distance, pull rope status) to roughly judge the "in / out of" safe state, unable to refine the identification of continuous stages such as "slight imbalance - high-risk imminent fall - fallen", let alone giving early warnings and hierarchical braking control according to the dynamic changes of human body postures.

[0007] In addition, the treadmill scenario has its own particularity: the user moves relative to the ground at a high speed on a long and narrow conveyor belt. Once there is a gait disorder or excessive forward / backward tilting of the torso, it may exceed the safe running area and be taken off the running belt within a very short time. Simply relying on traditional safety pull ropes, simple distance thresholds or post-fall identification of "fallen" often makes it difficult to intervene in time to avoid accidents. Therefore, there is an urgent need for a comprehensive technical solution that combines visual pose recognition and a treadmill control system. When it identifies that the user's pose is abnormal, the balance state deteriorates or the risk relative to the running belt edge gradually increases, it gives an early fall warning in advance, and automatically executes an emergency stop and necessary safety protection actions when it determines that a fall is inevitable or has already occurred, thereby improving the active safety level during the use of the treadmill. Based on this, it is necessary to propose a treadmill fall warning and emergency stop method and system based on pose recognition. Summary of the Invention

[0008] The technical objective of the present invention is to overcome the problems existing in the existing treadmill safety pull ropes, simple distance detection and post-fall identification solutions, such as strong dependence, only being able to passively stop after a fall, unable to perceive the imbalance risk in advance and having a high false alarm and miss rate. It provides a treadmill fall warning and emergency stop method and system based on pose recognition. By fusing and analyzing the human bone pose characteristics of the treadmill user and the running belt operation parameters, it can real-time evaluate the imbalance risk and predict the remaining time to reach the dangerous boundary, realizing early warning of fall risk, hierarchical deceleration and automatic emergency stop, thereby significantly enhancing the active safety protection ability during the use of the treadmill and reducing fall and secondary injury accidents.

[0009] To achieve the above technical objective, the technical solution of the present invention is as follows.

[0010] Firstly, this application proposes a treadmill fall warning and emergency stop method based on posture recognition, applicable to a treadmill including a running belt, drive motor, and electronic controller. The method comprises the following steps: S1, deploying an image acquisition device above or to the side of the running belt to acquire continuous human motion video images containing the treadmill's operating area; S2, performing human detection and posture recognition on the video images, extracting the coordinates of key skeletal points of the target user on the treadmill, and obtaining a posture feature sequence including trunk tilt angle, head height, hip and knee flexion / extension angles, foot landing state, and the position and velocity of the human body's center of gravity in the running belt coordinate system; S3, fusing the posture feature sequence with treadmill operating parameters (including running belt speed, acceleration, and effective safe zone length of the running belt) to construct a time-varying state vector, and inputting it into a pre-... The pre-trained imbalance risk assessment model outputs the user's imbalance risk index and the predicted remaining time to reach a dangerous state in real time; S4, the risk judgment threshold is adaptively adjusted based on the running belt speed, user stride frequency, and historical posture stability, and the imbalance risk index and predicted remaining time are jointly judged. When the imbalance risk index exceeds the preset warning threshold and the predicted remaining time is less than the first time threshold, a graded fall warning is triggered and the treadmill is controlled to reduce the running belt speed according to the preset deceleration curve; S5, the posture characteristics are continuously monitored in the warning state. When a sharp drop in the height of the human body's center of gravity, a trunk tilt angle exceeding the set angle range, and / or severe loss of skeletal key points that continues to exceed the second time threshold are detected, and the imbalance risk index exceeds the emergency threshold, a fall event is judged, an emergency stop command is issued to the electronic controller, and rapid deceleration to stop is executed.

[0011] As a further improvement, the posture recognition in step S2 includes: extracting at least two-dimensional or three-dimensional coordinates of the head, neck, torso, hip joint, knee joint, and ankle joint from each frame of video using a deep learning-based human skeletal keypoint detection network, and transforming the coordinates to the treadmill running belt coordinate system through camera calibration and running belt geometry parameters, for calculating the relative position and relative speed of the human body's center of gravity along the length of the running belt; recording posture features, risk index, and treadmill operating parameters throughout the warning and shutdown process, and marking and storing the fall or near-fall interval for subsequent model updates and safety strategy optimization. Specifically, by jointly determining the imbalance risk index based on posture recognition and the predicted remaining time, a warning is issued to the user before an actual fall occurs and progressive deceleration control is implemented, thereby achieving advanced identification of fall risk and emergency shutdown control within the specific constraint space of the treadmill.

[0012] As a further improvement, the imbalance risk index is determined at least based on the following formula:

[0013]

[0014] in, This represents the current position of the person's center of gravity along the length of the running belt. The center of the running belt safety zone;

[0015] The set length of the running belt safety zone; The velocity of the human body's center of gravity along the length of the running belt;

[0016] This is the current treadmill speed; The preset maximum calibration speed is used for normalization processing;

[0017] The current torso tilt angle; This is the preset maximum allowable torso tilt angle;

[0018] , , Let be the weight coefficient, and satisfy... .

[0019] As a further improvement, the predicted remaining time is estimated based at least on the distance and relative speed between the body's center of gravity and the trailing edge of the running belt, calculated according to the following formula:

[0020] in, This is the distance along the length of the running belt from the center of gravity of the person at the current moment. The speed of the running belt; The velocity of the human body's center of gravity along the length of the running belt; To prevent tiny positive numbers with a denominator of zero, if the predicted remaining time is less than a preset time threshold, it is determined that there is a risk of the device being thrown off the running belt.

[0021] As a further improvement, step S4 also includes using a multi-category posture or action recognition network to determine the current human body state, distinguishing non-falling actions such as squatting, bending over to tie shoelaces, and temporarily stopping from actual or near-falling states; only when the identified category is walking, jogging, or running and the imbalance risk index exceeds a preset threshold will the warning and emergency shutdown process be initiated, in order to reduce false alarms and unnecessary shutdowns.

[0022] Secondly, this application proposes a treadmill fall warning and emergency stop system based on posture recognition, comprising: a treadmill body for providing controllable speed running belt movement; an image acquisition device for acquiring continuous human motion video images including the running belt area of ​​the treadmill; a posture recognition and risk assessment processing unit electrically connected to the image acquisition device for extracting key skeletal points of the target user from the video images, calculating posture features and fusing treadmill operating parameters, and outputting an imbalance risk index and predicted remaining time in real time according to a preset imbalance risk assessment model; a warning and control unit electrically connected to the posture recognition and risk assessment processing unit and the treadmill drive controller for performing graded audible and visual warnings, running belt deceleration control, and emergency stop control based on the comparison results of the imbalance risk index and predicted remaining time with an adaptive threshold; and a data recording and communication unit for recording posture features, risk index, and treadmill operating parameters during a fall or near-fall event, and optionally uploading the relevant data to a host computer or cloud server via wired or wireless means, wherein the posture recognition and risk assessment processing unit and the warning and control unit are configured to execute the steps of the method described.

[0023] As a further improvement, the image acquisition device includes an RGB or depth camera mounted above the front side of the treadmill, with a top-down or tilt-down viewing angle. The camera's installation position and field of view are calibrated to cover the safe area of ​​the running belt and are fixedly connected to the treadmill frame to reduce the impact of machine vibration on posture recognition. The posture recognition and risk assessment processing unit has a built-in imbalance risk assessment model, which is a time-series-based deep learning model, including a one-dimensional convolutional neural network, a recurrent neural network, or a temporal Transformer, used to model posture features across multiple time steps, thereby predicting near-fall states in advance. The early warning and control unit is configured to: when the imbalance risk index is in the warning range, control the treadmill to gradually reduce the running belt speed according to a graded deceleration curve, and prompt the user with the current risk level through the display screen, indicator lights, and / or voice broadcast; when a fall event is determined, control the drive controller to cut off the drive motor power and trigger a mechanical or electrical braking mechanism to achieve an emergency stop, and optionally trigger a remote alarm or send a fall notification to a preset guardian terminal.

[0024] As a further improvement, a wearable sensor module or a treadmill-built-in pressure / inertial sensor module is also included. The sensor module is used to collect auxiliary data such as the user's plantar pressure distribution, triaxial acceleration, and angular velocity. The auxiliary data, together with the skeletal key point features obtained from posture recognition, are input into the posture recognition and risk assessment processing unit to improve the accuracy and robustness of the imbalance risk index and fall event determination.

[0025] Thirdly, this application proposes a computer device, the computer device including a processor, a graphics processing unit (GPU) and a memory, the memory storing a computer program, the computer program being executed by the processor and the GPU, causing the computer device to execute the posture recognition-based treadmill fall warning and emergency stop method.

[0026] Fourthly, this application proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a computer device, causes the computer device to perform a method for treadmill fall warning and emergency stop based on posture recognition.

[0027] This invention deeply integrates human skeletal posture recognition with treadmill operating parameters, achieving a shift from coarse detection of "whether on the running belt" to refined quantitative assessment of "imbalance trend - fall risk - dangerous time window." On one hand, it constructs an imbalance risk index using multi-dimensional posture features such as center of gravity position and speed, trunk tilt angle, lower limb joint flexion and extension state, and foot landing pattern. Combined with running belt speed, acceleration, and effective safety zone length, it calculates the predicted remaining time to the trailing edge of the running belt. This allows it to identify "high-risk near-fall state" before the user actually falls, issuing graded audio-visual / voice warnings in advance and automatically reducing the running belt speed according to a preset deceleration curve, effectively extending the time window for the user to self-correct their posture. On the other hand, it triggers emergency stop by using adaptive thresholds and multi-condition joint criteria (such as sudden drop in center of gravity height, abnormal trunk tilt, duration of loss of key posture points, etc.). This can quickly cut off the drive and execute braking when the user actually falls or is about to be thrown out of the running belt, significantly reducing the fall speed and impact intensity, and reducing secondary injuries such as abrasions and impacts. Meanwhile, this invention records fall and near-fall samples during the warning and shutdown process and uses them for model updates, enabling the system's risk assessment and control strategies to continuously adapt and optimize with the accumulation of usage data. This results in higher detection accuracy, lower false shutdown rate, and a smoother, more comfortable, and safer treadmill user experience, significantly improving the treadmill's active safety protection level and intelligence overall. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the treadmill fall warning and emergency stop system based on posture recognition according to the present invention;

[0029] Figure 2 This is a flowchart illustrating the treadmill fall warning and emergency stop method based on posture recognition according to the present invention. Detailed Implementation

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0031] (I) General Overview of the Implementation Examples

[0032] This embodiment provides a safety protection system deployed on a home or commercial treadmill. By installing a camera in front of or above the treadmill, it captures real-time video images of the user within the running belt area. A human posture recognition algorithm is run on an embedded computing unit or industrial control computer to obtain the user's skeletal key points and further calculate posture features such as the user's center of gravity position, center of gravity velocity, trunk tilt angle, lower limb joint flexion and extension states, and foot landing states. The system simultaneously acquires treadmill operating parameters (including running belt speed, acceleration, and running belt geometry), fuses them to form a temporal state vector, and outputs an imbalance risk index through an imbalance risk assessment model. and predicting the remaining time Then, by combining an adaptive threshold strategy, we can realize fall risk warning, graded deceleration control and emergency stop control, and record relevant data for subsequent model updates and strategy optimization.

[0033] The embodiments of this invention mainly consist of two parts: hardware system structure and method flow. The system structure can be referred to... Figure 1 For the functional block diagram and method steps, please refer to [the diagram / flow]. Figure 2 A flowchart.

[0034] (II) Terminology Explanation

[0035] To facilitate understanding of this invention, some terms used in the specification are explained first. These explanations are only for aiding understanding of the technical solutions and do not constitute a limitation on the claims:

[0036] 1. The running belt safety zone refers to a pre-defined section of length on the treadmill running belt, typically with a safe center point a certain distance from the front of the treadmill. Extend forward and backward by a certain length When a person's center of gravity is within this area, they are considered to be in a relatively safe position.

[0037] 2. Location of the body's center of gravity It refers to the projected position of the body's overall center of gravity along the length of the running belt, estimated based on key points of the human skeleton. It is defined in the running belt coordinate system and can be represented by numerical coordinates.

[0038] 3. Trunk angle The angle between the human torso and the vertical direction can be calculated by connecting key points such as the shoulder and hip joints. Forward tilt is positive, and backward tilt is negative. The unit can be angle or radians.

[0039] 4. Imbalance Risk Index It refers to a normalized risk metric that is weighted by multiple factors such as center of gravity offset, the difference between the center of gravity and the running belt speed, and the torso tilt angle. It is a comprehensive indicator used to characterize the degree of imbalance of the user at the current moment.

[0040] 5. Predict the remaining time It refers to the time required for the human body's center of gravity to reach the dangerous position at the rear edge of the running belt, based on the current distance between the human body's center of gravity and the trailing edge of the running belt, as well as the relative motion trend. It is used to characterize the time margin before the human body is thrown out by the running belt.

[0041] 6. Early Warning Threshold and Emergency Threshold: The early warning threshold refers to the index of imbalance risk. A set of lower thresholds is set to trigger audible and visual warnings and gradual deceleration; the emergency threshold refers to a higher threshold, which is used to trigger an emergency shutdown when multiple conditions are met.

[0042] 7. The near-fall state refers to a state in which one has not actually fallen but has already shown a clear tendency to lose balance, such as a rapid shift of the center of gravity backward, severe forward or backward leaning of the torso, and obvious gait disorder.

[0043] (II) Overall System Structure (corresponding) Figure 1 )

[0044] like Figure 1 As shown, the treadmill fall warning and emergency stop system based on posture recognition of the present invention mainly includes the following modules:

[0045] 1. The treadmill body includes the running belt, frame, drive motor, drive controller, and human-machine interface panel, and is the basic equipment for users to conduct running training. The drive controller can adjust the running belt speed, decelerate, and stop the machine according to external control signals.

[0046] 2. The image acquisition device is preferably an RGB camera, but a depth camera or an RGB-D camera can also be used. The camera is fixedly mounted on the crossbeam above the front side of the treadmill, ensuring its field of view covers the safe area of ​​the running belt. The camera is connected to the lower-level computing unit via a USB, MIPI, or Ethernet interface.

[0047] 3. The posture recognition and risk assessment processing unit can adopt an embedded ARM board, edge computing box, or industrial control computer. It internally deploys a deep learning inference engine to run the human detection network and skeletal keypoint detection network, calculate posture features based on keypoints, fuse treadmill data, and output the results. and The processing unit communicates with the drive controller via serial port, CAN bus, or Ethernet.

[0048] 4. The early warning and control unit can be integrated into the processing unit or implemented as a separate module, used to... and The comparison results with the adaptive threshold generate control strategies, including audible and visual warnings, voice prompts, graded deceleration control, and emergency stop commands, and the control commands are sent to the treadmill drive controller.

[0049] 5. The data recording and communication unit is used to record the postural characteristics, risk index curve, treadmill operating parameters and trigger operations during a fall or near-fall event, and can upload the data to a cloud server or host computer via Wi-Fi or Bluetooth for system maintenance, model updates and safety analysis.

[0050] The above modules are configured into a unified system via wired or wireless means. Those skilled in the art can select appropriate hardware devices and communication methods based on specific product requirements without affecting the essence of this invention.

[0051] (III) Method Implementation and Technical Route (corresponding) Figure 2 )

[0052] The main steps of the method of the present invention are as follows: Figure 2 As shown, steps S1 through S6 are included. The specific implementation methods of each step are explained in detail below.

[0053] S1: Video Acquisition and System Initialization

[0054] In this embodiment, after the system is powered on, it first initializes the camera, drive controller, and processing unit. The processing unit opens the video acquisition interface of the image acquisition device through the driver program and sets the resolution (e.g., ...). ), frame rate (e.g., 25-30 frames per second), and exposure mode (automatic or semi-automatic exposure).

[0055] To ensure the accuracy of attitude recognition, camera calibration and coordinate system transformation parameter calibration are required after camera installation. Camera calibration can be performed using a checkerboard calibration board, employing conventional camera calibration algorithms to obtain camera intrinsic parameters and distortion coefficients. Then, several physical calibration points are selected on the running belt (e.g., multiple locations at the leading, mid, and trailing edges), and their pixel coordinates are located in the camera image. A mapping relationship from the image coordinate system to the running belt plane coordinate system is established using planar homography or PnP calculation methods. Those skilled in the art can complete this process using commonly used open-source libraries.

[0056] During system initialization, the processing unit also reads parameters such as the current running belt speed, acceleration, and running belt length from the treadmill drive controller, and sets the initial safe zone center position. and safe zone length It loads the weight files of the pre-trained human detection model and skeletal keypoint detection model. Simultaneously, it initializes the imbalance risk assessment model, threshold parameters, and data recording cache.

[0057] When the user stands on the running belt and presses the start button, the treadmill begins operation. At this time, the processing unit begins acquiring video frames from the camera at fixed time intervals and simultaneously reads the latest running belt speed from the drive controller. Operating parameters such as acceleration are used as the basis for subsequent attitude analysis.

[0058] S2: Human Detection, Pose Recognition, and Pose Feature Extraction

[0059] This step is one of the core components of the invention, responsible for accurately extracting the target user's posture information from the video.

[0060] 1. The human target detection and tracking processing unit 30 first executes a human detection network, such as a lightweight detection network based on YOLO, CenterNet, etc., on each frame of input image, and outputs the human bounding box in the image. To improve efficiency and stability, a target tracking algorithm (such as Kalman filtering + Hungarian matching) can be used to track between adjacent frames to ensure that pose estimation is always performed on the target user on the treadmill.

[0061] 2. Skeletal Keypoint Detection: For each target human bounding box, a skeletal keypoint detection network is invoked, such as a multi-person pose estimation algorithm based on a deep convolutional network or Transformer, to output the two-dimensional coordinates of key points such as the head, neck, shoulders, elbows, wrists, hips, knees, and ankles. And confidence level. If necessary, a single-person pose network can be used to improve accuracy.

[0062] 3. Key Point Coordinate Mapping and Filtering: Using the camera calibration results completed in S1, the coordinates of key points in the image are mapped to three-dimensional or two-dimensional coordinates in the running belt plane coordinate system. ,in The direction is along the length of the running belt. The direction is horizontal. First-order or second-order low-pass filtering can be applied to the key point time series to reduce jitter and noise.

[0063] 4. Human Body Center of Gravity Calculation: Based on the human biomechanical model, mass weights are assigned to each key skeletal point or body segment, and the position of the center of gravity in the running belt coordinate system is calculated. In discrete form, it can be represented as a weighted average of multiple key points. Those skilled in the art can use existing publicly available methods for calculating the center of gravity of a human body.

[0064] 5. Center of gravity velocity and acceleration estimation: The velocity along the length of the running belt is obtained by taking the difference of the center of gravity position in the time dimension. The stability can be enhanced by using a sliding window linear fitting or Kalman filtering method, in relation to the acceleration signal.

[0065] 6. Trunk tilt angle and joint angle calculation: Using the line connecting the midpoints of the shoulder and hip joints on both sides as the trunk axis, the angle between this axis and the vertical direction is calculated in the running belt coordinate system. Simultaneously, the angle between the hip, knee, and ankle joints can be calculated to characterize the flexion and extension state of the lower limbs, which can be used to distinguish between normal running swing and abnormal bending, squatting, and other movements.

[0066] 7. Gait Status and Foot Landing Pattern: By analyzing the vertical height changes and speeds of both ankle joints, the landing / lifting status of the left and right feet in each frame can be identified, thereby obtaining cadence, stride length, and foot support pattern, providing a basis for subsequent risk assessment and threshold adaptation.

[0067] The above steps yielded , , Joint angles, step frequency, etc., constitute a multi-dimensional pose feature vector. In each sampling period, the processing unit appends the pose features of the current frame to a fixed-length time window to form a pose feature sequence, which serves as the input for step S3.

[0068] S3: Fusion of Posture Characteristics and Treadmill Parameters and Imbalance Risk Assessment

[0069] This step constructs a time-series state vector based on the posture feature sequence obtained from S2 and the treadmill operating parameters, and calculates the imbalance risk index using an imbalance risk assessment model. and predicting the remaining time .

[0070] 1. The state vector is constructed at the current time step. Select the most recent Construct a temporal state vector from the pose features of frames (e.g., within a 1-2 second time window) and the corresponding treadmill running parameters. In a simplified embodiment, Includes the following components:

[0071] Current and historical center of gravity ;

[0072] Current and historical center of gravity speed ;

[0073] Current and historical trunk inclination ;

[0074] Lower limb joint angles, cadence, left and right foot support patterns, etc.;

[0075] treadmill speed Acceleration, running belt length wait.

[0076] In more advanced embodiments, one-dimensional convolutional and recurrent neural networks can be introduced to automatically extract features.

[0077] 2. Imbalance Risk Index calculate

[0078] In one alternative embodiment, the imbalance risk index It can be calculated using the following analytical form:

[0079]

[0080] in:

[0081] This represents the current position of the person's center of gravity along the length of the running belt.

[0082] The center of the running belt safety zone;

[0083] The set length of the running belt safety zone;

[0084] The velocity of the human body's center of gravity along the length of the running belt;

[0085] This is the current treadmill speed;

[0086] The preset maximum calibration speed is used for normalization processing;

[0087] The current torso tilt angle;

[0088] This is the preset maximum allowable torso tilt angle;

[0089] , , Let be the weight coefficient, and satisfy... .

[0090] By setting up appropriately , , This allows for a balance between the contributions of center of gravity shift, velocity mismatch, and torso posture to overall risk.

[0091] In another embodiment, it can be Input is fed into a pre-trained deep learning model, such as a one-dimensional convolutional network, LSTM, or temporal Transformer, and the model outputs... During training, supervised learning can be conducted using labeled data from experimentally collected samples of normal running, near falls, and falls.

[0092] Predict the remaining time calculate

[0093] By combining the geometric position of the treadmill's rear edge and the trend of the center of gravity movement, the remaining time can be calculated and predicted. In one embodiment, the following formula may be used:

[0094]

[0095] in:

[0096] This is the distance along the length of the running belt from the center of gravity of the person at the current moment.

[0097] The speed of the running belt;

[0098] The velocity of the human body's center of gravity along the length of the running belt;

[0099] To prevent tiny positive numbers with a denominator of zero.

[0100] when At that time, the human body moves backward relative to the running belt. This represents the time required to reach the trailing edge at the current relative velocity; if Then the denominator is limit, It will take a larger value to avoid false alarms.

[0101] pass and These two indicators can simultaneously provide a quantitative characterization of the current state from two dimensions: "severity of imbalance" and "time margin," which is more precise and reliable than simply relying on location or a static threshold.

[0102] S4: Adaptive threshold setting and joint judgment, as well as graded early warning and deceleration

[0103] This step is based on dynamic risk assessment results, combined with individual user differences and exercise intensity, to set adaptive thresholds and implement graded warnings and gradual deceleration control.

[0104] 1. The historical stability and gait characteristic statistical processing unit statistically analyzes the user's gait characteristics during the initial stage of treadmill operation or multiple training cycles, under stable running conditions. By analyzing parameters such as distribution, center of gravity position swing range, step frequency, and stride length, personalized stable posture characteristics are obtained. The stability window can be calculated within this range. The mean and standard deviation are used for subsequent dynamic threshold adjustment.

[0105] 2. Adaptive adjustment of warning threshold and emergency threshold: In one embodiment, the warning threshold can be adjusted... and emergency threshold Set to a function related to the current treadmill speed, cadence, and historical posture stability. For example, when the treadmill speed is high and the user's cadence is significantly unstable, Lowering the threshold appropriately makes the system more sensitive; when there are many false alarms in historical operations, the threshold can be appropriately increased. This can be implemented using empirical formulas or linear mappings, or through a simple rule engine.

[0106] 3. The joint decision rule system performs a check on each sampling period. and The pose features from the previous step are then used for joint determination. Taking one embodiment as an example, the following rules can be adopted:

[0107] like or Greater than the first time threshold If the condition is stable or slightly unbalanced, no warning will be triggered.

[0108] like and If it is determined to be a high-risk near-fall state, a warning state will be entered;

[0109] If the fall detection conditions are met subsequently (see S5), an emergency shutdown will be triggered.

[0110] 4. Gradual Warning and Deceleration: When the warning state is activated, the warning and control unit 40 first reminds the user to "watch your step" or "please adjust your running posture" through the buzzer, indicator lights, or voice broadcast on the treadmill panel. At the same time, it sends a gradual deceleration command to the treadmill drive controller. Instead of an immediate stop, the running belt speed is gradually reduced according to a preset deceleration curve, for example, linearly reducing from the current speed to a lower safe speed within 2 to 5 seconds.

[0111] The parameters selected during the deceleration process can be based on Size and Adjusting the urgency level: When the risk is extremely high and the time margin is extremely short, a steeper deceleration curve can be selected; when the risk has just exceeded the threshold and the time margin is still acceptable, a slower deceleration curve can be adopted to reduce interference with the user's movement rhythm.

[0112] 5. False Alarm Suppression and Fault Tolerance Strategies: In the early warning state, the system continues to monitor attitude characteristics in real time. When a false alarm is detected... Rapidly descend to a safe range and rise to greater than If the torso tilt angle and cadence return to the normal range, the system considers the user to have successfully regained balance. The system can then terminate the warning state, stop further deceleration, and even allow a slow recovery to the user-set target speed, thereby avoiding frequent false alarms and unnecessary shutdowns.

[0113] Through the aforementioned adaptive threshold and joint judgment mechanism, this invention can balance false alarm rate and user experience while ensuring safety, and limit warning and deceleration actions to be performed only when there is a real risk of falling.

[0114] S5: Fall Event Detection and Emergency Shutdown Control

[0115] When the system is in a warning or high-risk state, a more rigorous assessment is needed to determine whether a fall has already occurred or is about to occur inevitably, in order to decide whether to execute an emergency shutdown.

[0116] 1. Detection of sudden drop in center of gravity height: This test detects changes in center of gravity height over a period of time. If a significant drop occurs within a short period of time (e.g., 0.5 seconds) and falls below a certain percentage (e.g., 60%) of the normal standing height, it is considered that a fall or slump may have occurred.

[0117] 2. Detection and monitoring of abnormal trunk tilt When the absolute value of the tilt continuously exceeds the preset range (e.g., leaning forward more than 40° or leaning backward more than 30°) for a certain period of time, it can be considered an abnormal posture. In addition, a comprehensive judgment should be made in conjunction with the direction of the center of gravity shift, such as leaning forward while the center of gravity moves forward, or leaning backward while the center of gravity moves backward towards the rear edge, etc.

[0118] 3. Keypoint Missing and Occlusion Detection: When a user falls, some key points (such as the head and hip joints) may be obscured and unable to be recognized by the detection network. The system can statistically analyze the confidence and number of keypoint detections across multiple consecutive frames. When the total number of keypoints significantly decreases and consistently exceeds a second time threshold, the system will detect the missing keypoints. This can be used as an auxiliary condition for determining a fall.

[0119] 4. Combined triggering conditions In one embodiment, the system combines the above conditions with , Combined, set up a fall detection strategy:

[0120] ;

[0121] Or the center of gravity has already crossed a certain distance beyond the rear edge of the safe zone;

[0122] A sudden drop in center of gravity or a prolonged exceedance of trunk angle;

[0123] And one of the conditions for missing key points is met.

[0124] When several of the above conditions are met simultaneously, it can be determined that a fall has occurred or that a fall is unavoidable. The system will immediately send an emergency stop command to the drive controller, triggering the braking mechanism to stop the treadmill in the shortest feasible time.

[0125] 5. Emergency Shutdown and Follow-up: In the event of an emergency shutdown, the control unit can simultaneously trigger an audible and visual alarm or send a notification to a pre-set monitoring terminal (for home or rehabilitation scenarios). For commercial gym scenarios, the corresponding treadmill number and alarm reason can be displayed on the management console, prompting staff to check the user's status.

[0126] S6: Fall / Near Fall Data Recording and Model Update

[0127] To ensure that system performance can continuously improve over time, this embodiment records data when warnings, near-fall or fall events occur, and uses this data as the basis for subsequent model training and threshold optimization.

[0128] 1. Data Recording Content: When the system determines that it has entered a warning state, is in a near-fall state, or a fall event has occurred, it will record the original posture feature sequence, treadmill operating parameters, and imbalance risk index within a certain time window before and after the current moment. Curve, Predicted Remaining Time The curve, along with the final control actions (whether to decelerate or stop urgently), are packaged and stored in local memory.

[0129] 2. The event labeling and classification system allows maintenance personnel to label events after they occur through a simple human-computer interaction interface. For example, the label can be "false alarm", "real fall", "near fall but not yet fallen", "user actively stopped", etc., which can be used for subsequent supervised or semi-supervised training of the model.

[0130] 3. Model and Threshold Optimization: In feasible implementations, the collected data can be uploaded to a cloud server, allowing researchers to retrain or fine-tune the imbalance risk assessment model using larger-scale training data. The new model weights are then distributed to the terminal devices. For analytical methods... The calculation scheme can also be adjusted based on the statistically obtained false alarm rate and false negative rate. , , And adjust each threshold.

[0131] (iv) Specific application example: Safety verification test for fall prevention on home treadmills

[0132] 1. Experimental objective:

[0133] The practical effects of the present invention, "Treadmill Fall Warning and Emergency Stop Method and System Based on Posture Recognition," in the following aspects are verified:

[0134] Is it possible to predict the remaining time before the user actually falls, using the imbalance risk index R(t) and other factors? To achieve early warning and deceleration;

[0135] Can it quickly and urgently stop the machine in the event of a fall or when a fall is unavoidable, thereby reducing the impact of the fall and the risk of secondary injury?

[0136] Compared with traditional safety rope solutions and single distance detection solutions, can it significantly reduce the incidence of falls and accidental machine stoppages, and improve the active safety and comfort of treadmill use?

[0137] 2. Test equipment and objects

[0138] Six home smart treadmills of the same model were selected, and each was equipped with one of three safety strategies:

[0139] Option A: Only a traditional magnetic safety rope is configured (Comparative Example 1);

[0140] Option B: Add a single distance sensor to Option A to detect when the distance between the user and the control panel exceeds the limit and then shut down the machine (Comparative Example 2).

[0141] Option C: Configure the fall warning and emergency stop system based on posture recognition of the present invention (example).

[0142] In Scheme C, the treadmill is equipped with a camera, processing unit, and early warning control module according to the specific implementation method in the instruction manual. The system parameters (center position of the safety zone, length of the safety zone, weight coefficient, etc.) have been set during the installation and debugging phase.

[0143] 3. The trial population will consist of 20 adult participants, including:

[0144] Ten healthy volunteers with some running experience;

[0145] Ten volunteers who usually exercise little or have little running experience.

[0146] All participants signed informed consent forms before the trial and were supervised by professionals.

[0147] 4. Test Environment: The test was conducted in an indoor gym environment with a flat, non-slip floor. A safety buffer zone was provided behind each treadmill, and soft mats were laid to ensure that the risk of injury was minimized even if the system did not stop in time.

[0148] 5. Test Condition Design

[0149] To simulate real-world risk scenarios as closely as possible, this example designed the following typical working conditions, with each subject completing the same experimental sequence on the treadmill under three different scenarios (crossover design to avoid order bias):

[0150] 5.1 Condition 1: Normal warm-up and steady-speed running

[0151] Set the treadmill speed to 6km / h~8km / h for 5 minutes;

[0152] Subjects were instructed to maintain a steady running pace and not to intentionally create any abnormal movements.

[0153] This operating condition is used to statistically analyze the false alarms and accidental shutdowns of each scheme under normal conditions, and to verify the stability and comfort of the system.

[0154] 5.2 Operating Condition 2: Slight Imbalance and Brief Retreat

[0155] At a running belt speed of 8 km / h, the subjects deliberately simulated slight imbalance movements such as "slightly missing a step, taking the wrong step, or briefly taking two steps backward" as instructed, but were able to regain their balance on their own in the end;

[0156] Each subject must simulate the event at least 5 times.

[0157] This operating condition is used to observe whether each solution can identify the "near fall" state and whether it will cause unnecessary emergency shutdowns.

[0158] 5.3 Operating Condition 3: Significant imbalance but self-correcting.

[0159] At a running belt speed of 10km / h, a professional coach guided the subject to simulate a relatively obvious imbalance under safe conditions, such as suddenly reducing stride length, rapidly shifting the center of gravity backward, and slightly leaning the torso forward.

[0160] Subjects were asked to try their best to regain their balance by increasing their pace or moving forward.

[0161] The key inspection points for this operating condition are:

[0162] Whether the present invention's scheme can issue an early warning and appropriately slow down when the imbalance risk index increases significantly and the predicted remaining time decreases, so as to buy time for the subject to restore balance;

[0163] Can the proportional scheme provide effective assistance without causing unnecessary machine downtime?

[0164] 5.4 Condition 4: Controlled Drop Simulation

[0165] Under the protection of safety supervisors, the speed of the running belt was gradually increased to 8km / h to 9km / h, and the subjects were guided to simulate the tendency to fall by "giving up their steps and no longer actively following the running belt with their feet";

[0166] Once the safety officer determines that a risk exists, immediate manual intervention is conducted to prevent actual danger, and only the system's reaction behavior is recorded. This working condition is mainly used to verify whether the solution of this invention can trigger an emergency stop and significantly reduce the displacement and speed of the running belt before it stops in a very short time when a fall or an unavoidable fall occurs.

[0167] 6. System behavior and observation results of Implementation Scheme C

[0168] 6.1 Normal Running Condition (Condition 1) During the normal warm-up and steady-speed running of all subjects, the present invention continuously monitored the center of gravity position, center of gravity speed, and trunk tilt angle through posture recognition, calculating the imbalance risk index and predicting the remaining time. When the posture was stable, the center of gravity fluctuated slightly around the center of the safe zone, the imbalance risk index remained at a low level, and the predicted remaining time was much greater than the set time threshold. The experimental records showed that:

[0169] No false warnings or false decelerations were caused by the system of this invention;

[0170] When individual subjects experience a brief moment of distraction and uncoordinated gait, causing a slight shift in their center of gravity, the imbalance risk index will briefly increase. However, due to the limited magnitude and short duration, and the fact that the remaining time is not urgent, the system does not enter an early warning state.

[0171] Compared with the "distance detection scheme" in Comparative Example 2, the scheme of the present invention hardly produces false shutdowns during normal running, while in Comparative Example 2, due to the subject's natural arm swing and occasional forward or backward movement, there are a few false shutdown events, which affect the experience.

[0172] 6.2 Slight Imbalance Condition (Condition 2) When the subject intentionally simulates "minor missteps, brief pauses, or taking two steps backward," the system of this invention identifies a significant shift in the center of gravity relative to the center of the running belt's safe zone, a mismatch between the center of gravity speed and the running belt speed, and a change in the torso angle, leading to a short-term increase in the imbalance risk index. However, as the subject quickly regains balance by accelerating forward and adjusting arm swing, the center of gravity returns to the safe zone, and the imbalance risk index decreases accordingly, with the predicted remaining time showing a recovery.

[0173] For brief and minor imbalances, in most cases the imbalance risk index does not continuously exceed the warning threshold or the predicted remaining time is not lower than the time threshold, so the system will not enter the warning state;

[0174] In cases of slight imbalance with a slightly larger range of motion, the system will trigger a short-term Level 1 warning, with a buzzer sounding and the running belt speed decreasing slightly by 0.5 to 1 km / h according to a preset curve. With the help of slight deceleration, the subject can more easily recover their gait. Subsequently, the imbalance risk index drops back to a safe range, and the system automatically cancels the warning without performing an emergency stop.

[0175] In contrast, Comparative Example 1 (safety rope only) did not respond actively at all when there was slight imbalance; Comparative Example 2, which only relied on the distance threshold, did not necessarily trigger an action when the subject stepped backward but had not yet clearly left the running belt, and its overall ability to "assist in correcting" slight imbalance was weak.

[0176] 6.3 Significant Imbalance but Self-Corrective Working Condition (Working Condition 3): The advantages of this invention are more pronounced when simulating significant imbalance at higher speeds. During the experiment, it was observed that when the subject suddenly reduced their stride or rapidly shifted their center of gravity backward, the system detected:

[0177] The center of gravity shifted rapidly, approaching the rear of the running belt;

[0178] The center of gravity speed is significantly lower than the running belt speed. , The significant increase causes the predicted remaining time to drop rapidly from a large value to near the set time threshold within a few seconds.

[0179] The increased forward tilt angle of the torso indicates a clear tendency to lunge forward.

[0180] At this point, the imbalance risk index increases and exceeds the warning threshold, while the predicted remaining time is less than the first time threshold. The system enters the warning phase, issuing a clear alert and smoothly reducing the running belt speed from 10 km / h to approximately 7 km / h within 2-3 seconds. Test records show:

[0181] In the vast majority of subjects, after receiving warnings and deceleration assistance, they were able to readjust their pace before falling, return their center of gravity to the safe zone, and stop the dangerous trend that could have developed into a fall.

[0182] If the system disables the warning and deceleration functions (as simulated in Comparative Examples 1 and 2), some inexperienced or physically weak subjects are more likely to be pushed further backward by the running belt under the same movements, requiring intervention from a safety officer to stabilize them.

[0183] This indicates that the present invention, through a graded early warning mechanism that combines an imbalance risk index with the predicted remaining time, significantly extends the time window for users to self-correct their posture and reduces the probability of entering a true fall state.

[0184] 6.4 Controlled Drop Simulation Condition (Condition 4)

[0185] In a controlled fall simulation, some participants followed instructions to "abandon following the running belt," allowing themselves to be gradually pulled backward. At this point, the system detected:

[0186] The center of gravity continues to shift towards the trailing edge of the running belt, and the predicted remaining time continues to decrease;

[0187] The increased difference between the center of gravity speed and the running belt speed caused the predicted remaining time to rapidly approach or even fall below the second time threshold.

[0188] Abnormal changes were observed in trunk tilt angle and the number of key points detected. Some subjects showed obvious forward or lateral tilting, and some key points of joints were lost.

[0189] When the above conditions are met simultaneously with the imbalance risk index exceeding the emergency threshold, the system immediately determines it as a fall event and sends an emergency stop command to the drive controller within a very short time. Records show:

[0190] The total time from entering a high-risk state to the treadmill coming to a complete stop was significantly shorter than that of comparative examples 1 and 2.

[0191] In the present invention, the distance that the user is "dragged" on the running belt is significantly shortened. In most cases, only a slight imbalance or one knee touching the ground occurs before stopping the machine. There is no situation where the user is continuously dragged by the high-speed running belt.

[0192] In Comparative Example 1, due to the reliance on the safety rope, there were instances where the subject fell and was dragged for tens of centimeters on the running belt before the safety rope broke.

[0193] In Comparative Example 2, when the subject's body was still within the range of the distance sensor during the fall, the system would not trigger a shutdown in a short period of time, and the dragging distance was significantly greater than that of the system of the present invention.

[0194] Based on a comprehensive assessment of the severity of the accident, the dragging distance, and the subjective discomfort experienced after the fall, the solution of this invention can significantly reduce the risk of impact and secondary injury when a fall occurs.

[0195] (V) Summary of Technical Effects

[0196] As can be seen from the above specific application examples, the treadmill fall warning and emergency stop method and system based on posture recognition of the present invention has the following comprehensive technical advantages compared with the traditional safety rope solution and the single distance detection solution:

[0197] In normal running scenarios, the posture recognition and risk assessment mechanism minimizes false alarms and accidental shutdowns, ensuring a continuous exercise experience for users.

[0198] In cases of slight imbalance and significant imbalance that can still be self-corrected, the system uses a combination of the imbalance risk index and the predicted remaining time to provide early warnings and gradual deceleration, effectively extending the time and space for users to self-correct and significantly reducing the probability of actual falls.

[0199] In controlled fall simulation, the solution of the present invention can quickly trigger an emergency stop when a fall occurs or a fall is unavoidable, shortening the drag distance of the user on the running belt and reducing the risk of fall impact and secondary injury.

[0200] Overall, this invention upgrades the "passive power-off after the fact" to an active safety solution of "pre-event risk prediction + process collaborative control + rapid braking after the fact", which significantly improves the inherent safety level and intelligence level of the treadmill during use.

[0201] Therefore, it can be seen that the technical solution described in this invention can achieve significant safety protection and user experience improvement in practical applications, and has good promotional value for both home and commercial treadmills.

[0202] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

Claims

1. A method for fall warning and emergency stop of a treadmill based on posture recognition, comprising the following steps: S1. Install an image acquisition device above or to the side of the running belt to acquire continuous human motion video images containing the running area of ​​the treadmill; S2. Perform human body detection and posture recognition on the video image, extract the coordinates of the key skeletal points of the target user on the treadmill, and obtain a posture feature sequence including torso tilt angle, head height, hip and knee flexion and extension angles, foot landing state, and the position and speed of the human body's center of gravity in the running belt coordinate system. S3. The posture feature sequence is fused with the treadmill operating parameters to construct a state vector that changes over time. This vector is then input into a pre-trained imbalance risk assessment model, and the user's imbalance risk index and the predicted remaining time to reach a dangerous state are output in real time. S4. Adaptively adjust the risk judgment threshold based on the running belt speed, user step frequency and historical posture stability, and jointly judge the imbalance risk index and the predicted remaining time. When the imbalance risk index exceeds the preset warning threshold and the predicted remaining time is less than the first time threshold, trigger the graded fall warning and control the treadmill to reduce the running belt speed according to the preset deceleration curve. S5. In the early warning state, continuously monitor posture characteristics. When a sudden drop in the height of the human body's center of gravity, a trunk tilt angle exceeding the set angle range and / or a serious loss of key skeletal points that continues to exceed the second time threshold, and the imbalance risk index exceeds the emergency threshold, it is determined to be a fall event. An emergency stop command is issued to the electronic controller and rapid deceleration is performed to stop the machine.

2. The treadmill fall warning and emergency stop method based on posture recognition according to claim 1, characterized in that: The posture recognition in step S2 includes: extracting at least two-dimensional or three-dimensional coordinates of the head, neck, torso, hip joint, knee joint, and ankle joint from each frame of video using a deep learning-based human skeletal keypoint detection network, and transforming the coordinates to the treadmill running belt coordinate system through camera calibration and running belt geometry parameters to calculate the relative position and relative speed of the human body's center of gravity along the length of the running belt; recording posture features, risk index, and treadmill operating parameters throughout the warning and shutdown process, and marking and storing the fall or near-fall interval for subsequent model updates and safety strategy optimization. Specifically, by jointly determining the imbalance risk index based on posture recognition and the predicted remaining time, a warning is issued to the user before an actual fall occurs and progressive deceleration control is implemented, thereby achieving advanced identification of fall risk and emergency shutdown control within the specific constraint space of the treadmill.

3. The method for fall warning and emergency stop of a treadmill based on posture recognition according to claim 1, characterized in that: The imbalance risk index is determined at least based on the following formula: ； in, This represents the current position of the person's center of gravity along the length of the running belt. The center of the running belt safety zone; The set length of the running belt safety zone; The velocity of the human body's center of gravity along the length of the running belt; This is the current treadmill speed; The preset maximum calibration speed is used for normalization processing; The current torso tilt angle; This is the preset maximum allowable torso tilt angle; , , Let be the weight coefficient, and satisfy... .

4. A treadmill fall warning and emergency stop method based on posture recognition according to claim 1 or 3, characterized in that: The predicted remaining time is estimated based at least on the distance and relative speed of the person's center of gravity relative to the trailing edge of the running belt, calculated according to the following formula: ； in, This is the distance along the length of the running belt from the center of gravity of the person at the current moment. The speed of the running belt; The velocity of the human body's center of gravity along the length of the running belt; To prevent tiny positive numbers with a denominator of zero, when the predicted remaining time is less than a preset time threshold, it is determined that there is a risk of the device being thrown out of the running belt.

5. A method for fall warning and emergency stop of a treadmill based on posture recognition according to claim 1, characterized in that: In step S4, the current human body state is determined by a multi-class posture or action recognition network, and the non-falling actions such as squatting, bending over to tie shoelaces, and temporarily stopping are distinguished from the actual fall or near fall state. The warning and emergency shutdown process is only initiated when the identified category is walking, jogging, or running and the imbalance risk index exceeds a preset threshold, in order to reduce false alarms and unnecessary shutdowns.

6. A treadmill fall warning and emergency stop system based on posture recognition, characterized in that, include: The treadmill body is used to provide controlled speed running belt motion; Image acquisition device for acquiring continuous human motion video images including the treadmill running belt area; The posture recognition and risk assessment processing unit is electrically connected to the image acquisition device. It is used to extract the target user's skeletal key points from the video image, calculate posture features and fuse treadmill running parameters, and output the imbalance risk index and predict the remaining time in real time according to the preset imbalance risk assessment model. The warning and control unit, electrically connected to the posture recognition and risk assessment processing unit and the treadmill drive controller, is used to perform graded audible and visual warnings, running belt deceleration control, and emergency stop control based on the comparison results of the imbalance risk index and the predicted remaining time with the adaptive threshold; the data recording and communication unit is used to record the posture characteristics, risk index, and treadmill operating parameters during a fall or near-fall event, and optionally upload the relevant data to a host computer or cloud server via wired or wireless means, wherein the posture recognition and risk assessment processing unit and the warning and control unit are configured to perform the method described in any one of claims 1 to 5.

7. A treadmill fall warning and emergency stop system based on posture recognition according to claim 6, characterized in that: The image acquisition device includes an RGB or depth camera mounted above the front of the treadmill, with a top-down or tilt-down viewing angle. The camera's installation position and field of view are calibrated to cover the safe area of ​​the running belt and are fixedly connected to the treadmill frame to reduce the impact of machine vibration on posture recognition. The posture recognition and risk assessment processing unit has a built-in imbalance risk assessment model, which is a time-series-based deep learning model, including a one-dimensional convolutional neural network, a recurrent neural network, or a temporal Transformer, used to model posture features across multiple time steps, thereby predicting near-fall states in advance. The early warning and control unit is configured to: when the imbalance risk index is in the warning range, control the treadmill to gradually reduce the running belt speed according to a graded deceleration curve, and prompt the user with the current risk level through the display screen, indicator lights, and / or voice broadcast; when a fall event is determined, control the drive controller to cut off the drive motor power and trigger a mechanical or electrical braking mechanism to achieve an emergency stop, and optionally trigger a remote alarm or send a fall notification to a preset guardian terminal.

8. A treadmill fall warning and emergency stop system based on posture recognition according to claim 6, characterized in that: It also includes a wearable sensor module or a treadmill-built-in pressure / inertial sensor module. The sensor module is used to collect auxiliary data including the user's plantar pressure distribution, triaxial acceleration, and angular velocity. The auxiliary data, together with the skeletal key point features obtained from posture recognition, is input into the posture recognition and risk assessment processing unit to improve the accuracy and robustness of the imbalance risk index and fall event determination.

9. A computer device, characterized in that, The computer device includes a processor, a graphics processing unit (GPU), and a memory, wherein the memory stores a computer program that, when executed by the processor and the GPU, causes the computer device to perform the method as described in any one of claims 1 to 5.

10. A computer-readable storage medium having a computer program stored thereon, the computer program, when executed by a computer device, causing the computer device to perform the method as described in any one of claims 1 to 5.

Images