Method for predicting fall risk with respect to a target object using a robot, and control device

Abstract

The present application relates to a kind of method for predicting fall risk about target object using robot, control device.The method includes: at robot, obtain multimodal data about target object, the multimodal data at least includes visual data, physiology data about target object;Based on the image frame, obtain the three-dimensional key point sequence of target object in continuous time window, so as to identify gait event based on the three-dimensional key point sequence and extract gait feature;Based on the gait feature, output gait risk representation data via gait stability score model;From the blood pressure data and heart rate data of the target object obtained, extract blood pressure and heart rate feature, so as to generate physiological risk variable based on the extracted key blood pressure and heart rate feature;And at least based on gait risk representation data, physiological risk variable, predict fall risk level about target object via multi-factor fusion model.

Description

Technical Field

[0001] Embodiments of the present invention relate to the field of robotics, and more specifically to a method and control device for predicting the fall risk of a target object using a robot. Background Technology

[0002] Traditional fall risk assessment methods typically rely on a single sensor. For example, they might rely solely on accelerometers in wearable devices to detect sudden posture changes or visual cameras to identify falling movements. Some mobile devices, such as smartphones, have attempted to integrate fall detection functionality. They use built-in accelerometers and gyroscopes to detect gait stability and alert the user when instability is detected. However, the core functionality of such single-modal solutions is largely focused on recognizing and alerting the user after a fall (e.g., fall event detection), failing to comprehensively capture precursory signals and lacking the ability to synchronously perceive early warning information, thus failing to provide early warning of fall risk. Furthermore, solutions based on built-in smartphone sensors rely on the phone's own motion signals, making them susceptible to being misled by simple shaking or vibrations, resulting in numerous false gait instability detections and false alarms, significantly reducing the reliability of risk assessments and user experience.

[0003] In summary, the shortcomings of traditional methods for assessing fall risk are at least that they are difficult to capture early warning signs that could lead to a fall, thus failing to provide early warnings of fall risk. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method and control device for predicting the fall risk of a target object using a robot. This method can not only identify the state after a fall occurs, but also predict the fall risk level in advance.

[0005] According to a first aspect of the present invention, a method for predicting fall risk of a target object using a robot is provided, comprising: acquiring multimodal data about the target object at the robot, the multimodal data including at least visual data and physiological data about the target object, the visual data including at least consecutive image frames about the target object, and the physiological data including blood pressure data and heart rate data of the target object; acquiring a three-dimensional keypoint sequence of the target object within a consecutive time window based on the image frames, so as to identify gait events and extract gait features based on the three-dimensional keypoint sequence; outputting gait risk characterization data based on the gait features via a gait stability scoring model; extracting blood pressure and heart rate features from the acquired blood pressure and heart rate data of the target object, so as to generate physiological risk variables based on the extracted key blood pressure and heart rate features; and predicting the fall risk level of the target object via a multi-factor fusion model, at least based on the gait risk characterization data and the physiological risk variables.

[0006] In some embodiments, obtaining a three-dimensional keypoint sequence of a target object within a continuous time window based on the image frames includes: inputting the image frames into a human detection model and outputting detection boxes about the target human body, the detection boxes indicating the position information of the target human body in the image and the confidence level of the human body category; generating a target trajectory of the target human body in a continuous time series based on the detection boxes of the continuous image frames; extracting image and depth information of the target human body region based on the detection boxes and the target trajectory for inputting into a pose estimation model to obtain a set of human keypoints, the keypoints including at least the head, shoulders, elbows, wrists, hips, knees, ankles, toes, and heels; and generating a three-dimensional keypoint sequence of the target object within a continuous time window based on the obtained set of human keypoints to identify gait events and extract gait features, the gait features including at least gait speed, stride length, stride frequency, stride width, bipedal support time ratio, gait asymmetry index, gait variation coefficient, trunk swing amplitude, center of mass trajectory, and center of pressure.

[0007] In some embodiments, predicting the fall risk level of a target object includes: calculating a gait and physiological interaction risk term; inputting gait risk representation data, physiological risk variables, and the gait and physiological interaction risk term into a multifactor fusion model to output a comprehensive fall risk score through weighted summation and a Sigmoid activation function; and determining the fall risk level of the target object based on the comprehensive fall risk score.

[0008] In some embodiments, the method for predicting the fall risk of a target object using a robot further includes at least one of the following: calculating an environmental risk variable based on visual recognition, depth ranging, and map construction results of the target object's environment; calculating a historical fall risk variable based on the target object's fall history; normalizing the raw scoring data of the acquired assessment scale for the target object to generate assessment scale risk variables; and calculating an activity behavior risk variable based on the target object's daily activity behavior patterns.

[0009] In some embodiments, predicting the fall risk level of a target individual includes: predicting the fall risk level of a target individual via a multi-factor fusion model based on gait risk characterization data, physiological risk variables, environmental risk variables, historical fall risk variables, assessment scale risk variables, activity behavior risk variables, and gait-physiological interaction risk items.

[0010] In some embodiments, blood pressure data includes systolic and diastolic blood pressure within an observation window, and heart rate data includes heart rate values ​​and RR interval sequences within an observation window. Generating physiological risk variables based on the extracted key blood pressure and heart rate features includes: calculating blood pressure level risk items, orthostatic hypotension risk items, abnormal heart rate risk items, heart rate variability risk items, and blood pressure and heart rate fluctuation risk items based on the target subject's systolic and diastolic blood pressure, heart rate values, and RR interval sequences; and multiplying the blood pressure level risk items, orthostatic hypotension risk items, abnormal heart rate risk items, heart rate variability risk items, and blood pressure and heart rate fluctuation risk items by their respective weighting coefficients, summing them, and then truncating the results to output the physiological risk variables.

[0011] In some embodiments, calculating the risk items for blood pressure level, orthostatic hypotension, abnormal heart rate, heart rate variability, and blood pressure and heart rate fluctuation includes: calculating systolic and diastolic blood pressure risk values ​​respectively using a piecewise mapping function based on the median systolic and diastolic blood pressure within the observation window to determine the blood pressure level risk item; detecting the change in blood pressure after the target subject changes from a sitting or lying position to a standing position to obtain the change in systolic and diastolic blood pressure, thereby calculating the orthostatic hypotension risk item; calculating the abnormal heart rate risk item based on the deviation of the median heart rate within the observation window from a preset heart rate reference range or the target subject's heart rate baseline; calculating a time-domain index based on the normal RR interval sequence within the observation window and calculating the heart rate variability risk item by comparing it with reference values ​​for healthy individuals; and calculating the blood pressure and heart rate fluctuation risk item based on the blood pressure coefficient of variation, heart rate coefficient of variation, and abnormal trends in blood pressure and heart rate within the observation window.

[0012] In some embodiments, the method of using a robot to predict the fall risk of a target object further includes: in response to determining that the target object has fallen, synchronizing the target object's location information and medical history information to a remote healthcare platform in real time; and outputting rescue instructions for the target object based on the disease type in the medical history information.

[0013] In some embodiments, outputting rescue instruction information for the target subject based on the disease type in the medical history information includes: in response to determining that the medical history information contains a history of heart disease, outputting information to instruct the subject to take emergency medication and adopt a semi-recumbent position; and in response to determining that the medical history information contains a history of stroke or that the current symptoms are consistent with signs of stroke, outputting information to instruct the subject to immediately call emergency services and avoid unnecessary movement.

[0014] In some embodiments, in response to determining that any of the following risk events have occurred, the weight of the physiological risk variables in the multifactor fusion model is increased: in response to the postural systolic blood pressure change value being greater than or equal to a first systolic blood pressure change threshold, or the postural diastolic blood pressure change value being greater than or equal to a first diastolic blood pressure change threshold, a risk event of orthostatic hypotension is determined; in response to determining that the time-domain index and frequency-domain high-frequency power are consistently lower than a preset heart rate threshold, and the low-frequency to high-frequency ratio is increased, a risk event of sympathetic over-excitation and vagal nerve activity is determined; in response to determining that the post-standing heart rate increase value is lower than a first predetermined heart rate threshold or the baseline heart rate fluctuation range relative to normal exceeds a predetermined value, a risk event of impaired baroreflex function is determined; and in response to determining that the blood pressure fluctuation amplitude during the morning peak exceeds a preset severe fluctuation threshold, or the systolic blood pressure coefficient of variation throughout the day exceeds a predetermined systolic blood pressure coefficient of variation threshold, and the heart rate variability parameter is consistently lower than a predetermined low heart rate variability threshold, a risk event is determined.

[0015] In some embodiments, the method of using a robot to predict the fall risk of a target object further includes: in response to determining that the fall risk level of the target object is low, the robot performs routine monitoring; in response to determining that the fall risk level of the target object is medium, the robot prompts to increase the monitoring frequency; and in response to determining that the fall risk level of the target object is high, the robot triggers a focus strategy.

[0016] According to a second aspect of the present invention, a control device is provided, comprising: at least one processing unit; and at least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions, when executed by the at least one processing unit, causing the control device to perform steps according to the method of the first aspect.

[0017] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0018] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements.

[0019] Figure 1 A schematic diagram of a system for predicting the fall risk of a target object using a robot is shown according to an embodiment of the present invention.

[0020] Figure 2 A flowchart of a method for predicting the fall risk of a target object using a robot, according to an embodiment of the present invention, is shown.

[0021] Figure 3 A flowchart of a method for extracting gait features according to an embodiment of the present invention is shown.

[0022] Figure 4 A flowchart illustrating a method for predicting the fall risk level of a target object according to an embodiment of the present invention is shown.

[0023] Figure 5 A flowchart of a method for adapting a monitoring strategy for a target object based on a fall risk level, according to an embodiment of the present invention, is shown.

[0024] Figure 6 A flowchart is shown for a method for increasing the weight of physiological risk variables in a multifactor fusion model according to an embodiment of the present invention. Detailed Implementation

[0025] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0026] The term "comprising" and its variations as used herein signify open inclusion, i.e., "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "at least partially based on". The terms "one example embodiment" and "one embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one additional embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0027] The following combination Figure 1 A system 100 is described for implementing a method of using a robot to predict the fall risk of a target object. Figure 1 A schematic diagram of a system 100 for implementing a method of predicting the fall risk of a target object using a robot, according to an embodiment of the present invention, is shown. Figure 1 As shown, system 100 includes robot 150, server 160, and network 170. In some embodiments, the control device 110 of robot 150 and server 160 interact with each other via network 170.

[0028] Regarding server 160, it may be deployed in a hospital information system, a server in a nursing home, or a cloud server. Server 160 is configured, for example, to interface with a hospital HIS system, electronic medical record system, follow-up system, and questionnaire system to obtain raw scoring data of assessment scales for the target subject and send it to the control device 110 of robot 150. Server 160 is also used to receive warning information from the robot. In some embodiments, the control device 110 of robot 150 independently performs fall risk prediction for the target subject and triggers a local alarm, and automatically synchronizes cached data to server 160.

[0029] Figure 1 A three-dimensional schematic diagram of robot 150 is shown. (As shown) Figure 1 As shown, the robot 150 includes at least: a control device 110, a torso 104, a head 102, a robotic arm, a supporting and moving part 113, a vision acquisition device, and an output device. Figure 1 (Not shown in the image). The output device includes at least a sound output device and a screen 134.

[0030] The torso 104 and head 102 are equipped with multiple data acquisition devices. These multiple data acquisition devices include, for example, at least a vision acquisition device. In some embodiments, the supporting movable part 113 is also equipped with a vision acquisition device. In some embodiments, the multiple data acquisition devices further include, for example, a blood oxygen sensor, a body temperature sensor, etc. In some embodiments, heart sounds are monitored via application software configured on a mobile terminal, and the detected heart sound monitoring data is transmitted to the control device 110 in the robot 150 via a cloud server. In some embodiments, the robot's head 102 is equipped with a millimeter-wave radar detection device, such that the millimeter-wave radar detection device is matched to the position of the target object's chest, thereby adapting to sense the chest rise and fall frequency of the target object. In some embodiments, the blood oxygen sensor measures, for example, using millimeter-wave radar. In other embodiments, the robot 150 is equipped with a pulse oximeter and its connection interface.

[0031] Regarding the robotic arm, it includes a left robotic arm 106 and a right robotic arm 108. One of the left and right robotic arms (e.g., left robotic arm 106) is equipped with a tactile hand, and the other of the left and right robotic arms (e.g., right robotic arm 108) is equipped with a blood pressure measuring device 132. In some embodiments, such as Figure 1 As shown, the left robotic arm 106 is flexibly connected to the haptic hand 112. The left robotic arm 106 and the right robotic arm 108 are each equipped with a six-dimensional force sensor, for example.

[0032] Regarding the blood pressure measuring device 132, it includes at least a blood pressure monitor, a blood pressure monitor control board, a blood pressure monitor microphone, a blood pressure monitor inflation pump, and a blood pressure monitor inflation solenoid valve. Figure 1 (Not shown separately). The blood pressure monitor control board is configured to control the blood pressure monitor inflation pump and the blood pressure monitor inflation solenoid valve to start inflation or deflation operations based on control commands from the control device 110.

[0033] The visual acquisition device is used to acquire visual data about a target object and its environment. The visual acquisition device may be one or more of the following: an RGB camera, an RGB-D depth camera, a binocular camera, a ToF depth camera, millimeter-wave radar, and lidar. In some embodiments, the visual acquisition device is configured on the robot's head. It should be understood that the visual data acquired by the visual acquisition device is provided to the control device 110 so that the control device 110, based on image frames in the visual data about the target object acquired by the robot, obtains a sequence of three-dimensional key points of the target object within a continuous time window, so as to identify gait events and extract gait features based on the three-dimensional key point sequence; and, based on the gait features, outputs gait risk characterization data via a gait stability scoring model.

[0034] A voice output device is configured to output voice prompts, such as, in response to determining that the medical history information includes a history of heart disease, outputting information instructing the user to take emergency medication and assume a semi-recumbent position; and in response to determining that the medical history information includes a history of stroke or that current symptoms are consistent with stroke signs, outputting information instructing the user to immediately call emergency services and avoid unnecessary movement. In some embodiments, the voice output device is also configured to output voice prompts based on output commands from a control device to prompt the target person to insert their arm into the blood pressure measuring device; or to output voice prompts to prompt the target person to raise their arm, extend their arm forward, and insert their arm into the blood pressure measuring device and remain relaxed; or to output voice prompts to prompt a nurse to assist the target person in inserting their arm into the blood pressure measuring device.

[0035] Regarding the support movement unit 113, it is used to move the robot based on control commands from the control device 110. For example, it drives the robot's support movement unit to move the robot to the environment where the target object is located. In some embodiments, the support movement unit 113 can also adjust the robot's chassis to a position suitable for the operation of the robotic arm based on control commands from the control device 110. The support movement unit 113 includes, for example, a movement component 130.

[0036] Regarding the control device 110, it is configured to acquire multimodal data about a target object, the multimodal data including at least visual data and physiological data about the target object, the visual data including at least consecutive image frames of the target object, and the physiological data including blood pressure and heart rate data of the target object; based on the image frames, acquire a three-dimensional keypoint sequence of the target object within a continuous time window, so as to identify gait events and extract gait features based on the three-dimensional keypoint sequence; based on the gait features, output gait risk characterization data via a gait stability scoring model. The control device 110 can also extract blood pressure and heart rate features from the acquired blood pressure and heart rate data of the target object, so as to generate physiological risk variables based on the extracted key blood pressure and heart rate features; and predict the fall risk level of the target object via a multi-factor fusion model, at least based on the gait risk characterization data and physiological risk variables.

[0037] Regarding the sound acquisition device, it is used to acquire the speech of the target object.

[0038] Regarding screen 134, it is at least used to present output results and indication information regarding the fall risk level of the target object. Control device 110 may have one or more processing units, including dedicated processing units such as GPUs, FPGAs, and ASICs, as well as general-purpose processing units such as CPUs. Additionally, one or more virtual machines may run on each control device 110.

[0039] The control device 110 includes, for example, a multimodal data acquisition module 1102, a gait feature extraction module 1104, a gait risk characterization data output module 1106, a physiological risk variable generation module 1108, and a fall risk level prediction module 1110. It should be understood that the multimodal data acquisition module 1102, the gait feature extraction module 1104, the gait risk characterization data output module 1106, the physiological risk variable generation module 1108, and the fall risk level prediction module 1110 are software modules running on one or more processing units included in the control device 110.

[0040] Regarding the multimodal data acquisition module 1102, it is used to acquire multimodal data about a target object at the robot. The multimodal data includes at least visual data and physiological data about the target object. The visual data includes at least continuous image frames about the target object, and the physiological data includes blood pressure data and heart rate data of the target object.

[0041] Regarding the gait feature extraction module 1104, it is used to obtain a sequence of three-dimensional key points of the target object within a continuous time window based on the image frame, so as to identify gait events and extract gait features based on the sequence of three-dimensional key points.

[0042] Regarding the gait risk characterization data output module 1106, it is used to output gait risk characterization data based on the gait characteristics and via a gait stability scoring model.

[0043] Regarding the physiological risk variable generation module 1108, it is used to extract key blood pressure and heart rate features from the acquired blood pressure and heart rate data of the target object, so as to generate physiological risk variables based on the extracted key blood pressure and heart rate features.

[0044] Regarding the fall risk level prediction module 1110, it is used to predict the fall risk level of a target object based on at least gait risk characterization data and physiological risk variables through a multi-factor fusion model.

[0045] Figure 2 A flowchart of a method 200 for predicting the fall risk level of a target object according to an embodiment of the present invention is shown. Method 200 may be derived from, for example... Figure 1 The control device 110 shown can be used to perform the operation, and it can also be used in... Figure 1 The method is performed at robot 150 as shown. It should be understood that method 200 may also include additional steps not shown and / or the steps shown may be omitted, and the scope of the invention is not limited in this respect.

[0046] In step 202, the control device 110 at the robot acquires multimodal data about the target object. The multimodal data includes at least visual data and physiological data about the target object. The visual data includes at least continuous image frames about the target object, and the physiological data includes blood pressure data and heart rate data of the target object.

[0047] In one implementation, the robot first acquires multimodal data via a data acquisition module. This multimodal data includes, for example, at least visual data and physiological data. In some embodiments, the multimodal data further includes: environmental data, fall history records, raw rating data from assessment scales, daily activity behavior pattern data, and historical disease data. It should be understood that the multimodal data is associated with timestamped data. For example, each frame of image, depth map, physiological data, and environmental data is timestamped 't' for synchronization of the multi-source data.

[0048] The visual data includes at least image frames of the target object. The visual data is acquired, for example, by the robot's camera or a camera system (e.g., one or more of an RGB-D camera, a binocular camera, and a ToF depth camera). In some embodiments, the robot also acquires data from millimeter-wave radar or lidar.

[0049] In some embodiments, visual data is represented as I_t ∈ R^(H×W×C), where I_t represents the image frame acquired at time t, H represents the image height, W represents the image width, and C represents the number of channels. When the robot's camera is a depth camera, the synchronously acquired depth map is expressed as D_t ∈ R^(H×W), where D_t represents the depth value corresponding to each pixel at time t.

[0050] The calibration data for the camera or camera system includes intrinsic parameters K and extrinsic parameters T_cw. Intrinsic parameter K includes focal lengths f_x and f_y and principal point coordinates c_x and c_y. Extrinsic parameter T_cw represents the transformation matrix between the camera coordinate system and the robot or world coordinate system.

[0051] The physiological data includes at least blood pressure and heart rate data. In some embodiments, the physiological data includes, but is not limited to: systolic blood pressure (SBP_t), diastolic blood pressure (DBP_t), heart rate (HR_t), heart rate variability (HRV_t), measurement posture (Posture_t), and measurement time (Time_t). In some embodiments, the physiological data is obtained via a blood pressure monitor, heart rate sensor, finger clip-on blood oxygen / heart rate module, wrist monitoring module, or wearable device that is communicatively connected to the robot.

[0052] Environmental data may include, for example, ground obstacles, slippery areas, ground friction coefficient, lighting, steps / slopes, passageway width, furniture edges, and high-risk area markings in restrooms / corridors.

[0053] In some embodiments, the data acquisition module preprocesses and synchronizes at least visual and physiological data to generate a multimodal data sequence: X_t = {I_t, D_t, K, T_cw, SBP_t, DBP_t, HR_t, HRV_t}. Here, I_t represents the image frame acquired at time t. D_t represents the depth value corresponding to each pixel at time t. T_cw represents the transformation matrix between the camera coordinate system and the robot or world coordinate system. SBP_t, DBP_t, HR_t, and HRV_t represent the systolic blood pressure, diastolic blood pressure, heart rate, and heart rate variability of the target object at time t, respectively.

[0054] In some embodiments, the data acquisition module preprocesses and synchronizes visual data, physiological data, environmental data, and historical disease data to generate a multimodal data sequence: X_t = {I_t, D_t, K, T_cw, SBP_t, DBP_t, HR_t, HRV_t, Env_t}, where Env_t represents the environmental risk data at time t.

[0055] In step 204, the control device 110 at the robot acquires a sequence of three-dimensional key points of the target object within a continuous time window based on the image frame, so as to identify gait events and extract gait features based on the sequence of three-dimensional key points.

[0056] A method for identifying gait events and extracting gait features based on the three-dimensional keypoint sequence includes, for example: inputting real-time acquired image frames into a human detection model and outputting detection boxes for the target human body, the detection boxes indicating the target human body's position information in the image and the human body category confidence; generating a target trajectory of the target human body in a continuous time series based on the detection boxes of consecutive image frames; extracting image and depth information of the target human body region based on the detection boxes and target trajectory for inputting into a pose estimation model to obtain a set of human keypoints, the keypoints including at least the head, shoulders, elbows, wrists, hips, knees, ankles, toes, and heels; generating a three-dimensional keypoint sequence of the target object within a continuous time window based on the obtained set of human keypoints, so as to identify gait events and extract gait feature vectors, the gait features including at least gait speed, stride length, stride frequency, stride width, bipedal support time ratio, gait asymmetry index, gait variation coefficient, trunk swing amplitude, center of mass trajectory, and center of pressure. The following will combine... Figure 3 The method 300 for extracting gait features is described in detail here and will not be repeated.

[0057] In some embodiments, the gait feature extraction module performs step 204, and its output gait feature vector is, for example, F_g = [v, StepLen, StepWidth, Cadence, DSR, ASI, CV_x, TrunkSway, CoM, COP,…]. Where v represents gait speed, StepLen represents gait frequency, StepWidth represents stride length, Cadence represents stride width, DSR represents bipedal support time ratio, ASI represents gait asymmetry index, CV_x represents gait variability coefficient, TrunkSway represents trunk swing amplitude, CoM represents center of mass trajectory, and COP represents center of pressure.

[0058] In step 206, the control device 110 at the robot outputs gait risk characterization data based on the gait characteristics and via a gait stability scoring model.

[0059] In some embodiments, the gait stability scoring module in the control device 110 at the robot is used to convert the above gait characteristics into a gait risk score G∈[0,1], where the closer G is to 1, the higher the gait risk.

[0060] First, the gait stability scoring module normalizes the gait features. The algorithm for this normalization is illustrated by the following formula (1).

[0061] z_m = (x_m - μ_m) / σ_m (1) In formula (1) above, x_m represents the m-th gait feature value. μ_m represents the reference mean of the gait feature. σ_m represents the reference standard deviation of the gait feature. z_m represents the gait feature after normalization.

[0062] Subsequently, the gait stability scoring module calculates gait risk characterization data based on the gait features after normalization. The following expression (2) exemplarily illustrates how the gait risk characterization data is calculated.

[0063] G = σ(a_0 + Σ_m a_m·z_m + Σ_m b_m·Δz_m) (2) In formula (2) above, σ represents the Sigmoid function. Δz_m represents the change in gait characteristics relative to the target object's own historical baseline. a_0 is the bias term, and a_m and b_m are weighting coefficients. The above method of calculating gait risk characterization data takes into account both the group reference standard and the individual baseline offset, and can more sensitively capture gait degradation trends.

[0064] In some embodiments, the present invention can predict gait risk representation data based on a temporal neural network (which may be a recurrent neural network RNN, a temporal convolutional network TCN, or a Transformer model) using gait feature sequences within a continuous time window. The following expressions (3) and (4) exemplarily illustrate another way of calculating gait risk representation data.

[0065] h_t = f_seq(F_{g,t-L+1}, …, F_{g,t}; θ_seq) (3) G = σ(W_g·h_t + b_g) (4) In the above expressions (3) and (4), F_{g,t} represents the gait feature vector at time t. f_seq represents the temporal model. θ_seq represents the temporal model parameters. L represents the time window length. h_t represents the hidden state output by the temporal model. W_g represents the weight matrix of the fully connected layer. b_g represents the bias term of the fully connected layer. The above method of calculating gait risk representation data can effectively capture the dynamic change pattern of gait features in the time dimension.

[0066] The following expression (5) exemplarily illustrates the training loss function representation of the model.

[0067] L_gait = -[y·log(G) + (1-y)·log(1-G)] + λ·||θ||_2^2 (5) In the above expression (5), y=1 represents a high-risk fall sample or a sample that has already fallen / is close to falling. y=0 represents a low-risk sample. G represents the gait risk representation data output by the model. θ represents all parameters of the model. λ represents the regularization coefficient, used to prevent the model from overfitting.

[0068] In step 208, the control device 110 at the robot extracts blood pressure and heart rate features from the acquired blood pressure and heart rate data of the target object in order to generate physiological risk variables based on the extracted blood pressure and heart rate features.

[0069] Regarding blood pressure data, this includes, for example, systolic and diastolic blood pressure within the observation window.

[0070] Regarding heart rate data, this includes, for example, heart rate values ​​within a window and RR interval sequences.

[0071] Regarding the calculation method of physiological risk variables, in some embodiments, it includes, for example, multiplying the risk items of blood pressure level, orthostatic hypotension, abnormal heart rate, heart rate variability, and blood pressure and heart rate fluctuation by their respective weighting coefficients, summing them, and then truncating the results to output physiological risk variables. The following expression (6) exemplarily illustrates the calculation method of physiological risk variables.

[0072] P = clip(η1·P_BP + η2·P_OH + η3·P_HR + η4·P_HRV + η5·P_var, 0, 1) (6) In the above expression (6), η1 to η5 are weighting coefficients. In some embodiments, exemplary values ​​may be η1=0.25, η2=0.25, η3=0.20, η4=0.15, and η5=0.15. The clip function is used to restrict the weighted sum to the interval [0,1]. P_BP represents the blood pressure level risk item, P_OH represents the orthostatic hypotension risk item, P_HR represents the heart rate abnormality risk item, P_HRV represents the heart rate variability risk item, and P_var represents the blood pressure and heart rate fluctuation risk item.

[0073] Methods for calculating the risk sub-item of blood pressure level include, for example, calculating the systolic blood pressure risk value and the diastolic blood pressure risk value respectively through a piecewise mapping function based on the median systolic blood pressure and the median diastolic blood pressure within the observation window, in order to determine the risk sub-item of blood pressure level.

[0074] The piecewise mapping functions include: the systolic blood pressure risk mapping function and the diastolic blood pressure risk mapping function. Table 1 below provides an example of how to calculate the systolic blood pressure risk value.

[0075] Table 1

[0076] Table 2 below illustrates how diastolic blood pressure risk values ​​are calculated.

[0077] Table 2

[0078] The blood pressure level risk sub-item, for example, takes the maximum of the systolic and diastolic blood pressure risk values. For instance, P_BP = max( _SBP(SBP_med), _DBP(DBP_med)). Here, SBP_med represents the median systolic blood pressure within the observation window, and DBP_med represents the median diastolic blood pressure. P_BP represents the risk factor for blood pressure levels.

[0079] Methods for calculating the risk of orthostatic hypotension include, for example, measuring the change in blood pressure of the target subject after changing from a sitting or lying position to a standing position, in order to obtain the change in systolic and diastolic blood pressure, thereby calculating the risk of orthostatic hypotension.

[0080] For example, when the target object changes from a sitting or lying position to a standing position, the changes in systolic and diastolic blood pressure are calculated. The following expressions (7) and (8) exemplarily illustrate how the changes in systolic and diastolic blood pressure are calculated.

[0081] ΔSBP = SBP_sit - SBP_stand (7) ΔDBP = DBP_sit - DBP_stand (8) In the above expressions (7) and (8), SBP_sit represents the systolic blood pressure in the sitting or lying position. SBP_stand represents the systolic blood pressure in the standing position. DBP_sit represents the diastolic blood pressure in the sitting or lying position. DBP_stand represents the diastolic blood pressure in the standing position.

[0082] The calculation method for the risk of orthostatic hypotension can be, for example, by using a continuous mapping. The following expression (9) exemplarily illustrates how the risk of orthostatic hypotension is calculated.

[0083] P_OH = clip(max(ΔSBP / 20, ΔDBP / 10), 0, 1) (9) In the above expression (9), P_OH represents the risk of orthostatic hypotension. When ΔSBP reaches 20 mmHg or ΔDBP reaches 10 mmHg, P_OH is set to 1. If blood pressure has not been measured within 3 minutes after standing, P_OH is set to the last valid value or the default value of 0.

[0084] The calculation method for the heart rate abnormality risk sub-item includes, for example, calculating the heart rate abnormality risk sub-item based on the degree of deviation between the median heart rate in the observation window and the preset heart rate reference range or the target object's heart rate baseline.

[0085] The following expression (10) exemplarily illustrates how the heart rate abnormality risk sub-item is calculated.

[0086] P_HR = clip(max(60 - HR_med, HR_med - 100, 0) / δ_HR, 0, 1) (10) In the above expression (10), P_HR represents the heart rate abnormality risk sub-item. HR_med represents the median resting heart rate within the observation window. δ_HR represents the heart rate deviation scale parameter (exemplarily taken as 20 bpm). In this formula, max(60 - HR_med, HR_med - 100, 0) represents the degree of deviation when the heart rate is below 60 bpm or above 100 bpm; the greater the deviation, the higher the risk. Optionally, if the target subject has a personalized baseline HR_base, it can be used instead of the general threshold for calculation. The following expression (11) exemplarily illustrates how to calculate the heart rate reference range.

[0087] P_HR = clip(|HR_med - HR_base| / δ_person, 0, 1) (11) In the above expression (11), δ_person represents the individualization deviation scale parameter.

[0088] Methods for calculating the risk of heart rate variability include, for example, calculating time-domain indices based on normal RR interval sequences within the observation window, and calculating the risk of heart rate variability by comparing them with reference values ​​for healthy individuals.

[0089] Regarding the RR interval, it is the time interval between two adjacent R waves in an electrocardiogram. It represents the duration of a complete cardiac cycle and is the basic raw data for calculating heart rate and heart rate variability (HRV). Heart rate variability reflects autonomic nervous system regulation. Decreased HRV is associated with an increased risk of falls. In one implementation, P_HRV is calculated based on time-domain indices SDNN (standard deviation of all normal RR intervals) and RMSSD (root mean square difference of adjacent RR intervals). The following expressions (12) and (13) exemplarily illustrate how SDNN and RMSSD are calculated.

[0090] SDNN = sqrt[(1 / N)·Σ_i(RR_i - RR_mean)^2] (12) RMSSD = sqrt[(1 / (N-1))·Σ_i(RR_{i+1} - RR_i)^2] (13) In the expressions (12) and (13) above, RR_i represents the i-th normal RR interval. RR_mean represents the mean of the RR intervals. N represents the total number of RR intervals.

[0091] The following expression (14) exemplarily illustrates how the heart rate variability risk sub-item is calculated.

[0092] P_HRV = clip(1 - min(SDNN / SDNN_ref, RMSSD / RMSSD_ref), 0, 1) (14) In the above expression (14), SDNN_ref and RMSSD_ref represent the time-domain reference values ​​for healthy elderly individuals (exemplary values: SDNN_ref = 100 ms, RMSSD_ref = 30 ms). When SDNN or RMSSD is much lower than the reference value, P_HRV is close to 1, indicating a decline in autonomic nervous system regulation and an increased risk of falls.

[0093] The calculation methods for the risk items of blood pressure and heart rate fluctuations include, for example, calculating the risk items of blood pressure and heart rate fluctuations based on the coefficient of variation of blood pressure and heart rate within the observation window, as well as the abnormal change trends of blood pressure and heart rate.

[0094] The blood pressure and heart rate variability risk sub-items are calculated based on the coefficient of variation and trend within the observation window. The following expressions (15), (16), and (17) exemplarily illustrate how the heart rate variability risk sub-item is calculated.

[0095] CV_SBP = std(SBP) / mean(SBP) (15) CV_HR = std(HR) / mean(HR) (16) P_var = clip(β1·CV_SBP + β2·CV_HR + β3·Trend_BP + β4·Trend_HR, 0,1) (17) In the above expressions (15), (16), and (17), CV_SBP represents the coefficient of variation of systolic blood pressure. CV_HR represents the coefficient of variation of heart rate. std represents the standard deviation calculation. mean represents the mean calculation. Trend_BP represents the abnormal upward or downward trend of the blood pressure curve (determined by the significance of the linear regression slope). Trend_HR represents the abnormal trend of the heart rate curve; β1 to β4 are weighting coefficients. When blood pressure or heart rate fluctuates drastically or shows a significant abnormal trend within the observation window, P_var increases accordingly.

[0096] In step 210, the control device 110 at the robot predicts the fall risk level of the target object based at least on gait risk characterization data and physiological risk variables via a multi-factor fusion model.

[0097] In some embodiments, the control device 110 at the robot predicts the fall risk level of the target object based at least on gait risk representation data G, physiological risk variable P, and a multi-factor fusion model. For example, in some embodiments, the control device 110 inputs the gait risk representation data G, physiological risk variable P, and gait-physiological interaction risk item I_GP into the multi-factor fusion model, and outputs a comprehensive fall risk score R through weighted summation and a sigmoid activation function, thereby determining the fall risk level of the target object. This achieves a multi-dimensional and comprehensive fall risk assessment.

[0098] In other embodiments, the control device 110 at the robot predicts the fall risk level of the target object via a multi-factor fusion model based on gait risk characterization data G, physiological risk variable P, environmental risk variable E, historical fall risk variable H, assessment scale risk variable Q, activity behavior risk variable A, and gait-physiological interaction risk item I_GP. When both gait risk characterization data and physiological risk variables increase simultaneously, the increase in fall risk often exhibits a superlinear relationship.

[0099] Regarding the gait-physiological interaction risk term I_GP, for example, when orthostatic hypotension is accompanied by gait instability, the actual fall risk is far higher than the simple sum of the two risks. To capture this interaction effect, this invention introduces the gait-physiological interaction risk term: I_GP = G · P. That is, the product of gait risk representation data and physiological risk variable. When both the gait risk representation data G and the physiological risk variable P are high, the value of the gait-physiological interaction risk term I_GP increases significantly, enabling the fusion model to provide a higher comprehensive risk score for target subjects with both gait abnormalities and physiological abnormalities, thereby more accurately reflecting the actual fall risk.

[0100] Training data for the model may include, for example, visual data collected by the robot, depth maps, 3D keypoint sequences, gait features, physiological parameters, and medical annotations. In some embodiments, training data may also include: environmental risk data, historical fall records, and raw score data from assessment scales.

[0101] In some embodiments, the present invention employs a multi-task learning approach for joint training, as exemplified by the following formula (18).

[0102] L_total = L_det + L_pose + L_gait + L_fusion + λ_reg·L_reg (18) In the above formula (18), L_det represents the human detection loss, L_pose represents the pose estimation loss, L_gait represents the gait stability score loss, L_fusion represents the fusion model loss, L_reg represents the regularization term, and λ_reg represents the regularization coefficient. The calculation method for the multimodal fusion risk loss is, for example: L_fusion = -[y·log(R) + (1-y)·log(1-R)]. Here, R represents the comprehensive fall risk score, with a value range of [0,1]; y represents the true risk label of the sample, where y=1 indicates that the target object has fallen or is in a high-risk state, and y=0 indicates that the target object has not fallen or is in a low-risk state; L_fusion represents the fusion risk loss function.

[0103] For multi-level risk classification tasks, cross-entropy loss can be used instead of binary cross-entropy loss; for scenarios where the recall rate of high-risk users needs to be focused on, Focal Loss can be used as the loss function to alleviate the problem of imbalance between positive and negative samples.

[0104] In some embodiments, regarding model updates, the server 160 of system 100 can periodically retrain the multi-factor fusion model based on newly collected data and send the updated model parameters to the control device 110 of robot 150. To protect user privacy, the system prioritizes uploading 3D keypoint sequence data, gait features, and risk summaries rather than raw video data; in one embodiment, federated learning or differential privacy methods are used for model updates to ensure that user data is not uploaded or only anonymized feature data is uploaded, thus meeting privacy protection requirements.

[0105] The comprehensive fall risk score R ranges from [0,1] and can also be converted to a percentage. In one embodiment, the present invention uses a three-level classification method to determine the fall risk level (i.e., low risk, medium risk, and high risk). If 0 ≤ R < 0.30 (or R_score < 30), the fall risk level for the target subject is determined to be low risk, and the robot prompts the user to maintain daily activities, regularly perform balance and strength training, and continue routine monitoring. If 0.30 ≤ R < 0.60 (or 30 ≤ R_score < 60), the fall risk level for the target subject is determined to be medium risk, and the robot prompts the user to increase the monitoring frequency, reminding the user to pay attention to getting up, turning, getting out of bed at night, and slippery environments, pushing risk warnings to caregivers or family members, and generating exercise and balance training suggestions. If 0.60 ≤ R ≤ 1.00 (60 ≤ R_score ≤ 100). If the fall risk level of the target individual is determined to be high, the robot will trigger a focused attention strategy, including proactive robot patrols, audio-visual alerts, notification of caregivers or family members, recommendations for age-friendly environmental modifications, and medical / rehabilitation assessments. If a fall, physiological abnormality, or prolonged stay occurs simultaneously, an emergency alarm will be triggered.

[0106] In the above scheme, the present invention can not only identify the immediate state after a fall event, but more importantly, it can output gait risk characterization data based on the extracted gait features, generate physiological risk variables based on the extracted blood pressure and heart rate features, and predict the fall risk level of the target object based on the gait risk characterization data and physiological risk variables through a multi-factor fusion model. Therefore, the present invention can build an advanced prediction mechanism to realize the pre-estimation and dynamic early warning of the probability of falls, thereby transforming passive response into active intervention.

[0107] Figure 3 A flowchart of a method 300 for extracting gait features according to an embodiment of the present invention is shown. Method 300 may be derived from, for example... Figure 1 The control device 110 shown can be used to perform the operation, and it can also be used in... Figure 1 The method is performed at robot 150 as shown. It should be understood that method 300 may also include additional steps not shown and / or the steps shown may be omitted, and the scope of the invention is not limited in this respect.

[0108] In step 302, the control device 110 at the robot inputs the image frame into the human detection model and outputs a detection box about the target human body. The detection box indicates the location information of the target human body in the image and the confidence level of the human body category.

[0109] In one implementation, the human body detection module in the control device 110 inputs real-time acquired image frames into the human body detection model and outputs detection boxes about the target human body.

[0110] The human detection module is constructed, for example, based on convolutional neural networks, YOLO series models, SSD models, Faster R-CNN models, DETR / Transformer detection models, or equivalent models. The input to the human detection module is an image frame I_t or RGB-D data {I_t, D_t}. The output of the human detection module is a detection box for one or more target human bodies. The following expression (19) exemplarily illustrates how the detection box for the target human body is expressed.

[0111] B_t = {b_t^1, b_t^2, …, b_t^N} (19) In the above expression (19), N represents the number of human targets detected in the current image frame. Let b_t^i represent the i-th detection box = (x_t^i, y_t^i, w_t^i, h_t^i, p_t^i), where x_t^i and y_t^i represent the coordinates of the top-left corner of the i-th detection box. w_t^i and h_t^i represent the width and height of the detection box. p_t^i represents the human body category confidence score. It should be understood that the detection box indicates the location information of the target human body in the image and the human body category confidence score.

[0112] The following expression (20) exemplarily illustrates how the human detection model is expressed as a function.

[0113] B_t = f_det(I_t; θ_det) (20) In the above expression (20), f_det represents the human detection model, θ_det represents the detection model parameters, and I_t represents the image frame. The following expression (21) exemplarily illustrates the algorithm for the loss function used to train the human detection model.

[0114] L_det = λ_cls·L_cls + λ_box·L_box + λ_obj·L_obj (21) In the above expression (21), L_cls represents the human body category classification loss. L_box represents the detection box regression loss. In some embodiments, it can be L1 loss, IoU loss, GIoU loss, DIoU loss, or a combination thereof. L_obj represents the target existence loss. λ_cls, λ_box, and λ_obj are the weight coefficients corresponding to each loss term.

[0115] In step 304, the control device 110 at the robot generates the target trajectory of the target human body in a continuous time series based on the detection boxes of continuous image frames.

[0116] In one embodiment, the target tracking module of the control device 110 acquires human detection results (i.e., detection boxes of target human bodies) and visual appearance features in multiple consecutive frames of images, calculates the matching cost between the currently detected target (i.e., the target object) and existing trajectories, assigns an identity number to the same target based on the matching result, and generates the target trajectory of the target object in a continuous time series. Thus, this invention maintains the identity number of the same object in consecutive image frames through the target tracking module, effectively avoiding identity drift or duplicate counting caused by occlusion, overlap, or temporary loss, and ensuring that subsequent gait analysis data comes from the same target.

[0117] The target tracking module takes as input the human detection results B_t from multiple consecutive frames and the visual appearance features e_t^i. The output of the target tracking module is the target trajectory of the target object in a continuous time series. The following expression (22) exemplarily illustrates how the target trajectory is expressed.

[0118] τ^i = {b_1^i, b_2^i, …, b_T^i, ID_i} (22) In the above expression (22), τ^i represents the trajectory sequence of the i-th target in the T-frame image. ID_i represents the identification number of the target.

[0119] Regarding the matching cost, it is used to associate the detected target with an existing trajectory. The following expression (23) exemplarily illustrates the functional expression of the matching cost.

[0120] C(i,j) = α_1·[1 - IoU(b_t^i, b̂_t^j)] + α_2·||e_t^i - ê_t^j||_2 + α_3·d_M(x_t^i, x̂_t^j) (23) In the above expression (23), IoU represents the intersection-union ratio of the detection boxes. e_t^i represents the visual appearance features of the current detected target. d_M represents the Mahalanobis distance, and α_1, α_2, and α_3 are the weight coefficients corresponding to each matching term.

[0121] In step 306, the control device 110 at the robot extracts the image and depth information of the target object's human body region based on the detection box and the target trajectory, so as to input the pose estimation model and obtain a set of human body key points, which include at least the head, shoulders, elbows, wrists, hips, knees, ankles, toes and heels.

[0122] Therefore, this invention can extract multiple dimensions at once, providing a rich data foundation for subsequent gait stability scoring.

[0123] In one implementation, the attitude estimation module in the control device 110 extracts the image and depth information of the human body region of the target object based on the detection box and the target trajectory, so as to input the attitude estimation model and obtain a set of human body key points.

[0124] The pose estimation model may be based on HRNet, OpenPose, BlazePose, ViTPose, or equivalent models. The input to the pose estimation model is the target human image region I_t[b_t^i], the depth map D_t, and the target trajectory τ^i. The output of the pose estimation model is a set of two-dimensional or three-dimensional human keypoints K_t^i = {k_t^{i,1}, …, k_t^{i,J}}, where J represents the number of keypoints. k_t^{i,j} = (u, v, z, s), where u and v represent pixel coordinates, z represents the depth value, and s represents the keypoint confidence. Keypoints include at least the head, shoulders, elbows, wrists, hips, knees, ankles, toes, and heels.

[0125] It should be understood that if a depth camera is used to acquire image frames, it is necessary to convert the two-dimensional keypoints into three-dimensional keypoint coordinates. The following expressions (24) to (26) exemplarily illustrate the functional expression of the matching cost.

[0126] X = (u - c_x)·z / f_x (24) Y = (v - c_y)·z / f_y; (25) Z = z (26) In the above expressions (24) to (26), c_x and c_y represent the principal point coordinates of the depth camera. f_x and f_y represent the focal length of the depth camera.

[0127] It should be understood that the present invention converts two-dimensional key points into true three-dimensional coordinates by combining depth maps and camera calibration parameters, making spatial parameters such as step length, step width, and centroid trajectory more accurate and unaffected by perspective distortion.

[0128] The following expression (27) provides an example of the functional expression for the attitude estimation model.

[0129] K_t^i = f_pose(I_t[b_t^i], D_t; θ_pose) (27) In the above expression (27), f_pose represents the pose estimation model, and θ_pose represents the model parameters. The following expression (28) exemplarily illustrates the algorithm for the loss function used to train the pose estimation model.

[0130] L_pose = λ_hm·L_heatmap + λ_kp·Σ_j||k_t^j - k_t^{j}||_1 + λ_b·L_bone (28) In the above expression (28), L_heatmap represents the heatmap loss. Σ_j||k_t^j - k_t^{j}||_1 represents the L1 loss of keypoint coordinates. L_bone represents the bone length consistency constraint loss. λ_hm, λ_kp, and λ_b represent the weight coefficients of the corresponding loss terms.

[0131] In step 308, the control device 110 at the robot generates a three-dimensional key point sequence of the target object within a continuous time window based on the obtained set of human key points, so as to identify gait events and extract gait features. The gait features include at least gait speed, stride length, stride frequency, stride width, bipedal support time ratio, gait asymmetry index, gait variation coefficient, trunk swing amplitude, center of mass trajectory, and center of pressure.

[0132] In some embodiments, the gait feature extraction module in the control device 110 generates a three-dimensional keypoint sequence of the target object within a continuous time window based on the obtained set of human keypoints, identifies gait events, and extracts gait feature vectors. In some embodiments, the gait feature extraction module outputs a gait feature vector, for example, F_g = [v, StepLen, StepWidth, Cadence, DSR, ASI, CV_x, TrunkSway, CoM, COP, …]. Wherein, v represents gait speed; StepLen represents gait frequency; StepWidth represents stride length; Cadence represents stride width; DSR represents bipedal support time ratio; ASI represents gait asymmetry index; CV_x represents gait variability coefficient; TrunkSway represents trunk swing amplitude; CoM represents center of mass trajectory; and COP represents center of pressure.

[0133] An algorithm for calculating walking speed is, for example, v = L_path / T_walk. Here, L_path represents the length of the walking path, and T_walk represents the walking time.

[0134] The algorithm for calculating the step size is, for example, StepLen = ||P_heel,left - P_heel,right||. StepLen indicates the distance between the left and right heels. P_heel,left represents the position of the left heel. P_heel,right represents the position of the right heel.

[0135] An example algorithm for calculating step frequency is: Cadence = 60·N_step / T_walk. Here, N_step represents the number of steps within the time window.

[0136] The algorithm for calculating stride width is, for example, StepWidth = |Y_ankle,left - Y_ankle,right|. StepWidth represents the lateral distance between the left and right ankle joints. Y_ankle,left represents the coordinates of the left ankle joint. Y_ankle,right represents the coordinates of the right ankle joint.

[0137] The algorithm for calculating the bipedal support time ratio is, for example, as follows: DSR = T_double / T_cycle. Here, T_double represents the bipedal support time, and T_cycle represents the time of one complete gait cycle.

[0138] The algorithm for calculating the gait asymmetry index is, for example, as follows: ASI = |X_L - X_R| / [(X_L + X_R) / 2]. Here, X_L and X_R represent the left and right gait parameters, respectively.

[0139] The algorithm for calculating the coefficient of variation of gait is, for example, CV_x = σ_x / μ_x. Here, σ_x represents the standard deviation of the gait parameters, and μ_x represents the mean of the gait parameters.

[0140] An example algorithm for calculating trunk swing amplitude is: TrunkSway = std(P_shoulder - P_pelvis). Here, P_shoulder represents the shoulder position, P_pelvis represents the pelvic position, and std represents the standard deviation.

[0141] An algorithm for calculating the centroid trajectory is, for example, CoM_t = Σ_j(m_j·X_t^j) / Σ_j(m_j). Here, m_j represents the mass of the j-th body segment, and X_t^j represents the position of the j-th body segment.

[0142] The algorithm for calculating the center of pressure is, for example, COP_x = Σ_k(p_k·x_k) / Σ_k(p_k). It should be understood that, when a plantar pressure sensor is configured, the center of pressure is calculated by weighting the pressure values ​​at each sensing point.

[0143] In the above scheme, the present invention can utilize the key point sequence within a continuous time window to capture the dynamic change trend of gait parameters, rather than relying solely on single frame or single step information. Therefore, it is more sensitive to instantaneous anomalies in gait.

[0144] Figure 4 A flowchart of a method 400 for extracting gait feature vectors according to an embodiment of the present invention is shown. Method 400 may be derived from, for example... Figure 1 The control device 110 shown can be used to perform the operation, and it can also be used in... Figure 1 The method is performed at robot 150 as shown. It should be understood that method 400 may also include additional steps not shown and / or the steps shown may be omitted, and the scope of the invention is not limited in this respect.

[0145] In step 402, environmental risk variables are calculated based on the visual recognition, depth ranging, and map construction results of the environment in which the target object is located.

[0146] In some embodiments, the environmental risk variable E is calculated by the robot based on visual recognition, depth ranging, and map construction results of the target object's home environment, community environment, or elderly care facility environment. The following expression (29) exemplarily illustrates how the environmental risk variable is calculated.

[0147] E = clip(e0+ e1·Obs + e2·Wet + e3·Light + e4·Slope + e5·Narrow +e6·Edge + e7·Area, 0, 1) (29) In the above expression (29), Obs represents obstacle density. Wet represents the risk of slippery ground. Light represents the risk of insufficient lighting. Slope represents the risk of ground slope or steps. Narrow represents the risk of narrow passage. Edge represents the risk of sharp furniture edges. Area represents high-risk area markers (such as restrooms, corridors, etc.). e0 to e7 are environmental risk weighting coefficients, and the clip function is used to limit the results to the interval [0,1].

[0148] In step 404, the historical fall risk variable is calculated based on the target object's fall history.

[0149] Regarding the historical fall risk variable H, it is calculated based on the target object's historical fall history. The following expression (30) exemplarily illustrates how the historical fall risk variable H is calculated. The following expression (31) exemplarily illustrates how the recent fall weight RecentWeight is calculated.

[0150] H = clip(1 - exp[-(r1·N_fall + r2·N_near + r3·Injury + r4·RecentWeight + r5·SceneRisk)], 0, 1) (30) RecentWeight = exp[-(T_now - T_lastfall) / τ] (31) In the expressions (30) and (31) above, N_fall represents the number of falls in the past 12 months. N_near represents the recent number of falls. Injury represents whether the fall caused injury. Recent represents the interval between the most recent fall and the current time. Scene represents the risk level of the scene in which the fall occurred. T_now represents the current time. T_lastfall represents the time when the most recent fall occurred. τ represents the time decay constant (an example value of 30 days). This time decay mechanism makes the more recent fall events have a greater impact on H, and can more sensitively reflect the current fall risk status of the target object.

[0151] In step 406, the raw scoring data of the obtained assessment scale for the target object are normalized in order to generate risk variables for the assessment scale.

[0152] Regarding the risk variable Q of the assessment scale, in some embodiments, it is calculated by the robot through human-computer interaction to obtain the self-assessment information of the target object.

[0153] In some embodiments, the robot uses the following three key questions (3KQ) for rapid initial screening: Question 1: Have you fallen in the past 12 months? Question 2: Do you feel unsteady when standing or walking? Question 3: Are you worried about falling? If the answer to each of the above questions is "yes", it is recorded as "1"; if the answer is "no", it is recorded as "0". The calculation method for the risk variable in the assessment scale is, for example: Q_3KQ = 0.4·q1 + 0.3·q2 + 0.3·q3. q1, q2, and q3 represent the records for questions 1 through 3, respectively. Q_3KQ represents the scoring data of the three key question risk assessment scales.

[0154] In other embodiments, the robot may acquire scores from a fall risk assessment scale and renormalize the original score data and weights of the fall risk assessment scale to obtain risk variables for the assessment scale.

[0155] Regarding fall risk assessment scales, these include, but are not limited to, the Morse Fall Risk Assessment Scale, the Fall Risk Self-Assessment Questionnaire for Older Adults (FRQ), the International Version of the Fall Effectiveness Scale (FES-I), the Activity-Specific Balance Confidence Scale (ABC), or the Timed Stand-Up-Walk Test (TUG) results. The raw scores of different scales are uniformly normalized to the [0,1] interval using the following formula (32): Q_scale = (S_scale - S_min) / (S_max - S_min) (32) In some embodiments, reverse normalization is used for scales (such as the ABC scale) where higher raw scores indicate lower risk. The following expression (33) exemplarily illustrates how reverse normalization is calculated.

[0156] Q_scale = 1 - (S_scale - S_min) / (S_max - S_min) (33) In the above expressions (32) and (33), S_scale represents the original scale score, and S_min and S_max represent the minimum and maximum possible scores in the scale, respectively.

[0157] The following expression (34) exemplifies how the risk variables of the assessment scale are calculated.

[0158] Q = clip(ρ1·Q_3KQ + ρ2·Q_Morse + ρ3·Q_FRQ + ρ4·Q_FESI + ρ5·Q_ABC +ρ6·Q_TUG, 0, 1) (34) In the above expression (34), ρ1 to ρ6 are the weight coefficients corresponding to each scale. Q_3KQ, Q_Morse, Q_FRQ, Q_FESI, Q_ABC, Q_TUG (0, 1) represent the normalized score data of the three key issue risk assessment scales, the Morse fall risk assessment scale, the Fall Risk Self-Assessment Questionnaire for Older Adults (FRQ), the International Version of the Fall Effectiveness Scale (FES-I), the Activity-Specific Balance Confidence Scale (ABC), or the Timed Stand-Up-Walk Test (TUG), respectively. When data for a certain scale is missing, the system re-normalizes the weights of the existing scales to ensure that the calculation of Q is not affected by missing data.

[0159] In step 408, activity behavior risk variables are calculated based on the target object's daily activity behavior patterns.

[0160] Regarding the activity behavior risk variable A, it is calculated based on the target object's daily activity behavior patterns. The following expression (35) exemplarily illustrates how the activity behavior risk variable A is calculated.

[0161] A = clip(a1·T_sitstand' + a2·TUG' + a3·Sedentary' + a4·NightOut' +a5·ActivityDrop' + a6·AbnormalStay' - a7·TrainingAdherence', 0, 1) (35) In the above expression (35), T_sitstand represents the time taken to stand up; TUG represents the time taken for the timed stand-up-walk test; Sedentary represents the duration of prolonged inactivity during the day; NightOut represents the duration of time spent out of bed at night; ActivityDrop represents the decrease in daily activity level compared to the individual's baseline; TrainingAdherence represents training adherence; and AbnormalStay represents the duration of abnormal stay in high-risk areas. Among these, variables marked with an apostrophe indicate the normalized variable values; a1 to a7 are the weight coefficients corresponding to each variable. Training adherence, as a negative indicator, indicates that a higher value appropriately reduces the risk of activity behavior, reflecting the positive role of consistent rehabilitation training in preventing and controlling the risk of falls.

[0162] In step 410, based on gait risk characterization data, physiological risk variables, environmental risk variables, historical fall risk variables, assessment scale risk variables, activity behavior risk variables, and interaction risk items, a multi-factor fusion model is used to predict the fall risk level of the target object.

[0163] For example, the aforementioned risk variables G, P, E, H, Q, A, and the gait-physiological interaction risk item I_GP are input into a multi-factor fusion model. Through weighted summation and a sigmoid activation function, a comprehensive fall risk score R is output, achieving a multi-dimensional and comprehensive fall risk assessment. It should be understood that in some embodiments, gait risk variables and physiological risk variables are given higher weights because they have the strongest direct correlation with fall occurrence. Environmental risk variables, historical risk variables, assessment scale risk variables, and activity-related behavioral risk variables are dynamically adjusted according to the specific application scenario. For example, if the target subject is in a hospital, the weights of the assessment scale risk variable Q and the historical fall risk variable H can be appropriately increased to fully utilize the assessment data from the medical institution. If the target subject is in a home environment, the weights of the environmental risk variable E and the activity-related behavioral risk variable A can be appropriately increased to reflect the significant impact of environmental factors and daily activity patterns on fall risk in the home environment. If the target subject is in a rehabilitation training scenario, the weights of gait risk G and activity-related behavioral risk A can be appropriately increased to more sensitively capture gait changes and improvements in mobility during the rehabilitation process. The following expression (36) exemplarily illustrates the objective function used for training.

[0164] min_{w,b} Σ_i{-y_i·log(R_i) - (1-y_i)·log(1-R_i)} + λ·||w||2 2 (36) In the above expression (36), R_i represents the comprehensive fall risk prediction value corresponding to the i-th sample, with a value range of [0,1]; y_i represents the true risk label of the i-th sample, where y_i=1 indicates that a fall has occurred or the sample belongs to a high-risk sample, and y_i=0 indicates that a fall has not occurred or the sample belongs to a low-risk sample; w_k represents the weight of the k-th risk variable, and w_k≥0; λ represents the regularization coefficient; ||w||2 2 This represents the L2 squared regularization term of the weight vector. The objective function uses cross-entropy loss to constrain the consistency between the risk prediction results and the true labels, and the regularization term avoids excessive weights, thus improving the model's generalization ability.

[0165] In some embodiments, the present invention can update individualized weights based on the target object's personal historical baseline. The calculation method of individualized weights is shown in, for example, the following formula (37).

[0166] w_user = (1 - λ_p)·w_global + λ_p·w_personal (37) In the above expression (37), w_global represents the global model weights trained based on group data, w_personal represents the personal model weights trained based on user personal data, and λ_p represents the individualized adjustment coefficient. This mechanism allows the model to benefit from the stability of training on large-scale data while also adapting to individual-specific change patterns.

[0167] Figure 5 A flowchart of a method 500 for adapting a monitoring strategy for a target object based on a fall risk level, according to an embodiment of the present invention, is shown. Method 500 may be derived from, for example... Figure 1 The control device 110 shown can be used to perform the operation, and it can also be used in... Figure 1 The method is performed at robot 150 as shown. It should be understood that method 500 may also include additional steps not shown and / or the steps shown may be omitted, and the scope of the invention is not limited in this respect.

[0168] In step 502, in response to the control device 110 determining that the fall risk level of the target object is low, the robot performs routine monitoring.

[0169] For example, if control device 110 determines that the fall risk score R satisfies: 0 ≤ R < 0.30 (or, R_score < 30), then the fall risk level for the target object is determined to be low, and the robot continues routine monitoring of the target object. In some embodiments, the robot also prompts the user to maintain daily activities and regularly perform balance and strength training.

[0170] Regarding methods for regularly engaging in balance and strength training, these include, for example, ensuring 150-300 minutes of moderate-intensity physical activity or 75-150 minutes of vigorous-intensity physical activity per week. Alternatively, following World Health Organization (WHO) recommendations, one can engage in at least 3 days of moderate-to-vigorous-intensity exercise per week, focusing on strengthening balance and muscle strength; and at least 2 days of moderate-to-vigorous-intensity exercise per week, specifically targeting major muscle groups. Combining these two types of exercise helps effectively maintain bone health and muscle function.

[0171] In step 504, in response to the control device determining that the fall risk level of the target object is medium risk, the robot prompts to increase the monitoring frequency.

[0172] If the control device 110 determines that the fall risk score R satisfies: 0.30 ≤ R < 0.60 (or, 30 ≤ R_score < 60), then the fall risk level for the target object is determined to be medium risk, and the robot increases the monitoring frequency for the target object. In some embodiments, the robot also prompts the user to pay attention to getting up, turning around, getting out of bed at night, and slippery environments, pushes risk alerts to caregivers or family members, and generates exercise and balance training suggestions.

[0173] Regarding recommendations for exercise and balance training, for example, this involves combining the actual physical function and activity level of the target individual (e.g., the elderly) to assist them in performing various types of functional movement training, such as standing up, squatting, and standing on one leg. These training movements are differentiated according to the target individual's joint range of motion, muscle strength, and balance ability. While ensuring training safety, external weight-bearing or measures to increase the range of motion or shorten the support time are selectively applied based on the target individual's real-time performance and tolerance, to achieve a step-by-step increase in training intensity. Furthermore, a fall risk assessment is conducted on the target individual at a preset period (preferably once a year), and the types of movements, load parameters, and difficulty coefficients in subsequent training programs are adjusted accordingly based on the assessment results, forming a closed-loop intervention management system.

[0174] In step 506, in response to the control device determining that the fall risk level of the target object is high, the robot triggers a focus strategy.

[0175] If the control device 110 determines that the fall risk score R meets the following condition: 0.60 ≤ R ≤ 1.00 (60 ≤ R_score ≤ 100), then the fall risk level for the target individual is determined to be high, and the robot triggers a focused attention strategy for the target individual. Such strategies may include: proactive robot patrols, audio-visual alerts, notifying caregivers or family members, suggesting age-friendly environmental modifications, and medical / rehabilitation assessments.

[0176] If the control device 110 determines that a fall, physiological abnormality, or prolonged stay has occurred simultaneously, an emergency alarm will be triggered.

[0177] In some embodiments, in response to the control device 110 determining that a target object has fallen, the device synchronizes the target object's location information and medical history information to a remote healthcare platform in real time; and outputs rescue instructions for the target object based on the disease type in the medical history information. For example, if the medical history information includes a history of heart disease, the device outputs instructions to take emergency medication and adopt a semi-recumbent position; if the medical history information includes a history of stroke or the current symptoms are consistent with stroke signs, the device outputs instructions to immediately call emergency services and avoid unnecessary movement. Before emergency personnel arrive at the scene, the robot terminal issues instructions on-site or pushes preliminary treatment suggestions to community volunteers to avoid blindly lifting or moving the object, which could lead to secondary injury. This assists rescue personnel in implementing safe and targeted treatment.

[0178] Regarding age-friendly environmental modifications and medical / rehabilitation assessments, for example, based on individualized assessments of the target group (e.g., the elderly), a multi-factor synergistic intervention strategy is implemented, encompassing one or more combinations of exercise training, nutritional support, medication adjustments, environmental modifications, and psychological counseling. In some embodiments, the target subjects are closely followed up for 30 to 90 days after the initial intervention. The follow-up interval can be set to once every 30, 60 or 90 days, depending on the condition of the target subjects, preferably once every 30 days. During the follow-up, the focus is on monitoring changes in the target subjects' fall risk, overall health status, progress of functional recovery and occurrence of adverse risk events. Based on the follow-up results, subsequent preventive and therapeutic measures are adjusted in a timely manner to form an adaptive closed-loop management.

[0179] In some embodiments, the multifactorial synergistic intervention strategy includes, for example, seven intervention modules: physical activity and balance training, pharmacological intervention, cardiovascular intervention, telemedicine and technology intervention, living environment intervention, intervention for vestibular disorders, pain and muscle strength, and fall fear intervention. Table 3 below illustratively illustrates the specific recommendations for each intervention module and their corresponding levels of evidence.

[0180] Table 3

[0181] Figure 6 A flowchart of a method 600 for increasing the weights of physiological risk variables in a multifactor fusion model according to an embodiment of the present invention is shown. Method 600 may be derived from, for example... Figure 1 The control device 110 shown can be used to perform the operation, and it can also be used in... Figure 1 The method is performed at robot 150 as shown. It should be understood that method 600 may also include additional steps not shown and / or the steps shown may be omitted, and the scope of the invention is not limited in this respect.

[0182] In step 602, it is determined whether any of the multiple risk events has occurred.

[0183] In some embodiments, continuous blood pressure and heart rate time series curves of the target object are acquired in real time or periodically, and key blood pressure and heart rate features for fall risk prediction are extracted from them. The key blood pressure and heart rate features include, but are not limited to: postural blood pressure variation values, blood pressure variability indicators, time-domain and frequency-domain indicators of heart rate variability, baseline heart rate fluctuation amplitude, and postural heart rate response values.

[0184] Regarding multiple risk events, these include, for example, at least: risk events of orthostatic hypotension, risk events of sympathetic over-excitation and vagal nerve activity, risk events of impaired baroreflex function, and risk events of heart rate variability.

[0185] In step 604, in response to determining that any risk event has occurred, the weights of the physiological risk variables in the multifactor fusion model are increased.

[0186] In some embodiments, in response to the postural systolic blood pressure change value being greater than or equal to a first systolic blood pressure change threshold, or the postural diastolic blood pressure change value being greater than or equal to a first diastolic blood pressure change threshold, a risk event of orthostatic hypotension is determined, and the weight of physiological risk variables in the multifactor fusion model is increased.

[0187] The threshold for the first change in systolic blood pressure is, for example, 20 mmHg.

[0188] The threshold for the first diastolic blood pressure change is, for example, 10 mmHg.

[0189] It should be understood that when using blood pressure data to assist in assessing fall risk, the magnitude of the decrease in postural systolic or diastolic blood pressure is the core criterion. Specifically, if, within 3 minutes of a target individual transitioning from a lying or sitting position to a standing position, systolic blood pressure decreases by 20 mmHg or more, or diastolic blood pressure decreases by 10 mmHg or more, it is considered orthostatic hypotension, at which point the risk of fall is significantly increased (preferably, the risk increment is approximately 25%). Furthermore, the faster the rate of blood pressure decrease and the greater the degree of frailty in the target individual, the higher the weighting coefficient for the increased fall risk in the predictive model needs to be.

[0190] In some embodiments, in response to determining that a risk event exists where the time-domain index and the high-frequency power in the frequency domain are consistently below a preset heart rate threshold, and the ratio of low frequency to high frequency increases, the weight of the physiological risk variable in the multifactor fusion model is increased.

[0191] When using heart rate data to aid in judgment, heart rate variability (HRV) is used as a strong early warning signal. Among them, persistently low time-domain indicators SDNN, persistently low high-frequency power (HF) in the frequency domain, and an increased low-frequency to high-frequency power ratio (LF / HF) all suggest excessive sympathetic nerve excitation and weakened vagal nerve activity, which are characteristic manifestations of increased risk. Therefore, it is necessary to increase the weight of physiological risk variables in the multi-factor fusion model.

[0192] In some embodiments, in response to determining that a risk event of impaired pressure reflex function exists when the heart rate increase after standing is lower than a first predetermined heart rate threshold or the baseline heart rate fluctuates more than a predetermined value relative to normal, it is necessary to increase the weight of physiological risk variables in the multifactor fusion model.

[0193] Regarding the preset value, it is, for example, 5 times / minute.

[0194] For example, a sluggish heart rate response after a change in body position (e.g., insufficient heart rate increase after standing), or a baseline heart rate that fluctuates by more than 5 beats per minute from the daily baseline, also suggests impaired pressure reflex function. In such cases, the weights of physiological risk variables in the multifactor fusion model should be increased or the risk level should be raised accordingly.

[0195] In some embodiments, in response to determining that the blood pressure fluctuation amplitude during the morning peak period exceeds a preset severe fluctuation threshold, or that the systolic blood pressure variability coefficient exceeds a predetermined systolic blood pressure variability coefficient threshold throughout the day, and that the heart rate variability parameter is consistently below a predetermined low heart rate variability threshold, it is determined that a heart rate variability risk event exists, and the weight of the physiological risk variables in the multifactor fusion model needs to be increased.

[0196] For example, in actual monitoring, the following early warning logic is set: when blood pressure fluctuates drastically during the morning peak period (preferably from 6:00 to 9:00 AM), or when the coefficient of variation of systolic blood pressure exceeds the normal range and increases abnormally throughout the day, and heart rate variability continues to decline, the subject's physiological risk is determined to be increased. In this case, the weight of the physiological risk variable in the multi-factor fusion fall risk prediction model is adaptively increased.

[0197] The various processes and procedures described above, such as methods 200 to 600, can be executed at a computing device. This computing device may include, for example, at least one processor (at least one graphics processor and at least one central processing unit); and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor. In some embodiments, methods 200 to 600 may be implemented as a computer software program or program product tangibly contained in a machine-readable medium. In some embodiments, part or all of the computer program may be loaded and / or installed on the computing device via read-only memory (ROM) and / or a communication unit. When the computer program is loaded into random-access memory (RAM) and executed by the GPU and CPU, one or more actions of methods 200 to 600 described above can be performed.

[0198] This invention can be a method, apparatus, system, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention. The computer-readable storage medium may be a tangible device capable of holding and storing instructions used by an instruction execution device. The computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof.

[0199] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network, to an external computer or external storage device. Various aspects of the invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0200] These computer-readable program instructions can be provided to the central processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the central processing unit of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of a flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of a flowchart and / or block diagram.

[0201] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0202] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps loaded in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.

[0203] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors.

Claims

1. A method for predicting the fall risk of a target object using a robot, comprising: At the robot, multimodal data about the target object is acquired. The multimodal data includes at least visual data and physiological data about the target object. The visual data includes at least continuous image frames about the target object, and the physiological data includes the target object's blood pressure data and heart rate data. Based on the image frames, a three-dimensional key point sequence of the target object within a continuous time window is obtained, so as to identify gait events and extract gait features based on the three-dimensional key point sequence; Based on the gait characteristics, gait risk characterization data is output via a gait stability scoring model; Key blood pressure and heart rate features are extracted from the blood pressure and heart rate data of the target subjects to generate physiological risk variables based on the extracted key blood pressure and heart rate features; as well as Based at least on gait risk characterization data and physiological risk variables, a multi-factor fusion model can be used to predict the fall risk level of the target subject.

2. The method according to claim 1, characterized in that, Based on the image frame, obtaining the sequence of 3D key points of the target object within a continuous time window includes: An image frame is input into a human detection model, which outputs a detection box about the target human body. The detection box indicates the location information of the target human body in the image and the confidence level of the human body category. Based on the detection bounding boxes of continuous image frames, the target trajectory of the target human body in a continuous time series is generated. Based on the detection box and target trajectory, the image and depth information of the human body region of the target object are extracted for input into the pose estimation model, thereby obtaining a set of human body key points, which include at least the head, shoulders, elbows, wrists, hips, knees, ankles, toes and heels; Based on the obtained set of human body key points, a three-dimensional key point sequence of the target object within a continuous time window is generated to identify gait events and extract gait features. The gait features include at least gait speed, stride length, stride frequency, stride width, bipedal support time ratio, gait asymmetry index, gait variation coefficient, trunk swing amplitude, center of mass trajectory, and center of pressure.

3. The method according to claim 1, characterized in that, The predicted fall risk level for the target object includes: Calculate the risk factors of gait-physiological interaction; Gait risk representation data, physiological risk variables, and gait-physiological interaction risk items are input into a multi-factor fusion model to output a comprehensive fall risk score through weighted summation and a sigmoid activation function; and Based on a comprehensive fall risk score, the fall risk level of the target group is determined.

4. The method according to claim 3, characterized in that, It also includes at least one of the following: Based on the visual recognition, depth ranging, and map construction results of the target object's environment, environmental risk variables are calculated; Calculate historical fall risk variables based on the target object's fall history; The raw scoring data of the assessment scales for the target subjects were normalized in order to generate risk variables for the assessment scales. as well as Based on the target group's daily activity behavior patterns, calculate activity behavior risk variables.

5. The method according to claim 4, characterized in that, The predicted fall risk level for the target object includes: Based on gait risk characterization data, physiological risk variables, environmental risk variables, historical fall risk variables, assessment scale risk variables, activity behavior risk variables, and gait-physiological interaction risk items, a multi-factor fusion model is used to predict the fall risk level of the target subject.

6. The method according to claim 4, characterized in that, Blood pressure data includes systolic and diastolic blood pressure within the observation window, and heart rate data includes heart rate values ​​and RR interval sequences within the observation window. Physiological risk variables are generated based on the extracted key blood pressure and heart rate features, including: Based on the target individual's systolic blood pressure, diastolic blood pressure, heart rate, and RR interval series, risk sub-items for blood pressure level, orthostatic hypotension, abnormal heart rate, heart rate variability, and blood pressure and heart rate fluctuations were calculated; and The risk items for blood pressure level, orthostatic hypotension, abnormal heart rate, heart rate variability, and blood pressure and heart rate fluctuation are multiplied by their respective weight coefficients, summed, and truncated to output physiological risk variables.

7. The method according to claim 6, characterized in that, The calculation of risk sub-items for blood pressure level, orthostatic hypotension, heart rate abnormalities, heart rate variability, and blood pressure and heart rate fluctuations includes: Based on the median systolic and diastolic blood pressure within the observation window, the systolic and diastolic blood pressure risk values ​​are calculated separately using a piecewise mapping function to determine the risk sub-items of blood pressure levels. The measurement of blood pressure changes in the target subject after changing from a sitting or lying position to a standing position is used to obtain the changes in systolic and diastolic blood pressure, thereby calculating the risk sub-item of orthostatic hypotension. Based on the degree of deviation between the median heart rate in the observation window and the preset heart rate reference range or the target subject's heart rate baseline, calculate the heart rate abnormality risk sub-item; Time-domain indices were calculated based on normal RR interval sequences within the observation window, and heart rate variability risk was calculated by comparing them with reference values ​​for healthy individuals; and Based on the blood pressure coefficient of variation, heart rate coefficient of variation, and abnormal trends in blood pressure and heart rate within the observation window, the risk sub-items of blood pressure and heart rate fluctuations are calculated.

8. The method according to claim 1, characterized in that, Also includes: In response to the determination that a target has fallen, the target's location information and medical history information are synchronized to the remote medical care platform in real time. as well as Based on the disease type in the medical history information, output rescue instructions for the target person.

9. The method according to claim 8, characterized in that, Based on the disease type in the medical history information, the output of rescue instructions for the target individual includes: In response to determining that the medical history information includes a history of heart disease, information is output to instruct the administration of emergency medication and the adoption of a semi-recumbent position; and In response to determining that the medical history information includes a history of stroke or that the current symptoms are consistent with signs of stroke, the system outputs information instructing the user to immediately call emergency services and avoid unnecessary movement.

10. The method according to claim 1, characterized in that, Also includes: In response to the determination of any of the following risk events, the weights of physiological risk variables in the multifactor fusion model will be increased: In response to the postural systolic blood pressure change being greater than or equal to a first systolic blood pressure change threshold, or the postural diastolic blood pressure change being greater than or equal to a first diastolic blood pressure change threshold, a risk event of orthostatic hypotension is determined; In response to the determination that the time-domain indicators and high-frequency power are consistently below the preset heart rate threshold, and the ratio of low frequency to high frequency is increased, a risk event of sympathetic overexcitation and vagal nerve activity is identified. In response to determining that the heart rate increase after standing is lower than a first predetermined heart rate threshold or that the baseline heart rate fluctuates more than a predetermined value relative to normal, a risk event of impaired baroreflex function is identified. as well as In response to determining that the blood pressure fluctuation amplitude during the morning peak period exceeds a preset severe fluctuation threshold, or that the systolic blood pressure coefficient of variation throughout the day exceeds a predetermined systolic blood pressure coefficient of variation threshold, and that the heart rate variability parameter remains below a predetermined low heart rate variability threshold, a risk event is determined to exist.

11. The method according to claim 1, characterized in that, Also includes: In response to determining that the fall risk level of the target object is low, the robot performs routine monitoring of the target object; In response to determining that the fall risk level of the target object is medium risk, the robot increases the monitoring frequency of the target object; as well as In response to determining that the fall risk level of the target object is high, the robot triggers a focused attention strategy for the target object.

12. A control device, comprising: At least one processing unit; At least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions, when executed by the at least one processing unit, causing the control device to perform the steps of the method according to any one of claims 1 to 11.

Images