Health care robot accompanying action intelligent optimization method and system assisted by AI vision

Abstract

The application discloses an AI vision-assisted health-care robot accompanying action intelligent optimization method and system, relates to the AI vision technical field, and comprises the following steps: collecting real-time image data of a target user through a tracking visual sensor arranged on a health-care robot, extracting gait coordination characteristic of the target user, inputting the gait coordination characteristic into a pre-trained lightweight stable state recognition model, outputting the center of gravity stability level of the target user from the lightweight stable state recognition model, matching a target accompanying instruction from a pre-set accompanying instruction set according to the center of gravity stability level, issuing the target accompanying instruction to the health-care robot for execution, and continuously acquiring real-time image data after execution to update the gait coordination characteristic. The technical problem that the existing health-care robot is not continuous in perceiving the gait state of a user and is not comprehensive in evaluating coordination during dynamic accompanying is solved, and the accompanying action response is lagged.

Description

Technical Field

[0001] This application relates to the field of AI vision technology, specifically to an AI vision-assisted intelligent optimization method and system for the movements of elderly care robots. Background Technology

[0002] With the continued growth in demand for elderly care services, the traditional human care model faces challenges such as manpower shortages, slow response times, and low service standardization. Against this backdrop, elderly care robots with autonomous perception and decision-making capabilities have become an important technological path to improve the quality of elderly care.

[0003] However, in the dynamic companionship process, existing health and wellness robots are prone to shaky, blurry, or target-losing images captured by visual sensors because both the user and the robot are in motion. This makes it difficult for the robot to obtain stable and reliable information about the user's walking status. At the same time, the robot's judgment on whether the user's walking is stable is often not timely enough. It often only reacts after the user has already shown obvious swaying, and cannot adjust its own position or movements in advance to provide support, making it difficult to achieve real-time coordination with the user's movement behavior.

[0004] Specifically, current mainstream technologies mostly rely on fixed-view cameras or wearable sensors to obtain information about the user's movement status. Fixed-view cameras are easily obstructed and have difficulty continuously tracking the target user's lower limb area, resulting in unstable gait feature extraction. While wearable sensors can provide more accurate motion data, they have drawbacks such as poor wearing compliance, low comfort during long-term use, and high equipment maintenance costs, which limit their widespread application in home-based health and wellness scenarios.

[0005] Furthermore, in terms of gait analysis, existing methods typically focus only on single-dimensional gait parameters, such as stride length or cadence, lacking comprehensive modeling of the key stability indicator of lower limb coordination. Gait coordination not only reflects the symmetry and rhythm of a user's walking, but is also the core basis for predicting fall risk. If the phase difference and variability of lower limb gait cannot be accurately quantified, it will be difficult to effectively assess the user's center of gravity stability, thereby affecting the timeliness and suitability of care strategies. Summary of the Invention

[0006] This application provides an AI-assisted intelligent optimization method and system for the companionship movements of elderly care robots, which solves the technical problem of discontinuous perception of user gait status and incomplete coordination assessment in existing elderly care robots during dynamic companionship, resulting in delayed response of companionship movements.

[0007] The technical solution to the above-mentioned technical problems in this application is as follows:

[0008] In a first aspect, this application provides an AI vision-assisted intelligent optimization method for the movements of a care robot companion, the method comprising:

[0009] Real-time image data of the target user is collected by a tracking visual sensor installed on the health and wellness robot;

[0010] Gait coordination features of the target user are extracted from the real-time image data, and the gait coordination features include at least the gait phase difference of the two lower limbs and the gait period variability;

[0011] The gait coordination features are input into a pre-trained lightweight stable state recognition model, which then outputs the center of gravity stability level of the target user.

[0012] Based on the center of gravity stability level, a target care instruction is matched from a preset care instruction set. The target care instruction is used to adjust at least one of the following: the movement speed, movement direction, or handrail height of the health care robot.

[0013] The target care instruction is sent to the health care robot for execution, and after execution, the robot continues to acquire the real-time image data to update the gait coordination characteristics.

[0014] Secondly, this application provides an AI vision-assisted intelligent optimization system for the movements of a care robot, including:

[0015] The image data acquisition module is used to acquire real-time image data of the target user through a tracking vision sensor installed on the health and wellness robot.

[0016] A gait coordination judgment module is used to extract the gait coordination features of the target user from the real-time image data. The gait coordination features include at least the gait phase difference of the two lower limbs and the gait period variability.

[0017] The recognition model training module is used to input the gait coordination features into a pre-trained lightweight stable state recognition model, and the lightweight stable state recognition model outputs the center of gravity stability level of the target user.

[0018] The care instruction matching module is used to match a target care instruction from a preset care instruction set according to the center of gravity stability level. The target care instruction is used to adjust at least one of the following: the moving speed, the moving direction, or the handrail height of the health care robot.

[0019] The care instruction execution module is used to send the target care instruction to the health care robot for execution, and continue to acquire the real-time image data after execution to update the gait coordination features.

[0020] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0021] This application provides an AI-assisted intelligent optimization method and system for the companion movement of a rehabilitation robot. First, it utilizes a tracking visual sensor mounted on the rehabilitation robot to achieve continuous, unobstructed tracking of the target user's lower limbs, avoiding the burden of wearing wearable devices and the challenges to compliance. Second, it extracts the gait phase difference and gait period variability of both lower limbs from real-time image data to construct multi-dimensional gait coordination features, improving the early detection sensitivity of center of gravity instability. Third, it employs a two-level lightweight classification architecture combined with a gating mechanism for a stable state recognition model, ensuring high discrimination accuracy while maintaining computational efficiency. Finally, it dynamically matches and executes appropriate companion commands based on the center of gravity stability level, enabling the robot to actively adjust its movement speed, direction, or handrail height.

[0022] Through the above technical solution, this application realizes a closed-loop optimization mechanism from gait perception and status recognition to caregiving response, achieving coordinated response with the user's gait behavior, thereby effectively reducing the risk of falls and improving the intelligence, personalization and safety of caregiving services. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the intelligent optimization method for AI vision-assisted care robot companion actions provided in the embodiments of this application;

[0025] Figure 2 This is a flowchart illustrating the process of obtaining the gait coordination characteristics of a target user in the AI ​​vision-assisted intelligent optimization method for the companion action of a health care robot provided in this application embodiment;

[0026] Figure 3 This is a schematic diagram of the structure of the AI ​​vision-assisted elderly care robot companion action intelligent optimization system provided in the embodiments of this application.

[0027] The components represented by each number in the attached diagram are explained below:

[0028] Image data acquisition module 11, gait coordination judgment module 12, recognition model training module 13, caregiver instruction matching module 14, and caregiver instruction execution module 15. Detailed Implementation

[0029] This application provides an AI-assisted intelligent optimization method and system for the companionship movements of elderly care robots, which addresses the technical problem of discontinuous perception of user gait status and incomplete coordination assessment in existing elderly care robots during dynamic companionship, resulting in delayed response of companionship movements.

[0030] Example 1, as Figure 1 As shown in the embodiments of this application, an intelligent optimization method for the actions of a healthcare robot with AI vision assistance is provided, including:

[0031] S10: Real-time image data of the target user is collected through a tracking vision sensor installed on the health and wellness robot;

[0032] In this embodiment, the tracking vision sensor adopts an active target tracking mechanism, which adjusts the pan-tilt angle and focal length of the camera in real time based on the lower limb region characteristics of the target user, ensuring that the key joints of the lower limbs are always completely preserved in the center of the image field of view during the user's walking process.

[0033] Specifically, step S10 in the method includes:

[0034] The tracking vision sensor includes an image acquisition unit and an attitude adjustment unit, wherein the image acquisition unit is used to acquire images, and the attitude adjustment unit is used to adjust the pitch angle, yaw angle, and focal length of the image acquisition unit.

[0035] The tracking vision sensor detects the target user and identifies the target user's lower limb area;

[0036] Based on the position and size of the lower limb region in the image, the pitch angle, yaw angle and focal length of the image acquisition unit are adjusted by the posture adjustment unit to keep the lower limb region within the preset area of ​​the image acquisition screen, and to keep the lower limb region intact and clear in the image;

[0037] During the movement of the target user, the tracking vision sensor continuously acquires real-time image data containing the lower limb region, and the real-time image data includes multiple frames of images arranged in chronological order.

[0038] In this embodiment, the health and wellness robot is equipped with a tracking vision sensor. This sensor is mounted on the robot body and can move synchronously with the robot to continuously track the target user and identify the target user's lower limb area. The tracking vision sensor includes an image acquisition unit and a posture adjustment unit. The image acquisition unit is used to acquire images and can use an RGB camera or a depth camera. The frame rate is usually set to 30 frames per second to ensure the temporal resolution of gait analysis. The posture adjustment unit is used to adjust the pitch angle, yaw angle, and focal length of the image acquisition unit so that the sensor can adaptively follow the movement of the target user.

[0039] During the target user's movement, the tracking vision sensor continuously adjusts the posture of the image acquisition unit through the posture adjustment unit, so that the lower limb area is kept within a preset area of ​​the image acquisition screen, such as the central area of ​​the screen, occupying 40% to 60% of the screen area, and the lower limb area remains complete and clear in the image. The real-time image data includes multiple frames of images arranged in chronological order, and each frame contains a complete and clear lower limb area, providing a data basis for subsequent gait analysis.

[0040] The tracking visual sensor detects the target user and identifies the target user's lower limb region, including:

[0041] Human skeletal key point detection is performed on the first frame image acquired by the image acquisition unit to obtain the skeletal key point positions of the target user. The skeletal key point positions include at least the positions of the left hip joint, right hip joint, left knee joint, right knee joint, left ankle joint, and right ankle joint.

[0042] Based on the location of the key skeletal points, a bounding box is determined that surrounds the left and right lower limbs, and the bounding box is used as the lower limb region;

[0043] In subsequent frames, a lightweight target tracking algorithm is used to track the lower limb region, and the position and size of the lower limb region are updated according to the tracked position.

[0044] When the tracking confidence of the lightweight target tracking algorithm is lower than the preset confidence threshold, the human skeleton key point detection is re-performed. After the location of the skeleton key point is re-acquired and the lower limb region is re-determined, the tracking of the lower limb region is resumed.

[0045] In this embodiment, the tracking vision sensor first detects the target user and identifies the lower limb region of the target user. This process uses a human skeletal key point detection algorithm, such as the OpenPose algorithm, to process the first frame image and obtain the skeletal key point positions of the target user. The skeletal key points include at least the positions of the left hip joint, right hip joint, left knee joint, right knee joint, left ankle joint, and right ankle joint. Based on the key point positions, a bounding box surrounding the left and right lower limbs is determined, and this bounding box is taken as the lower limb region. Specifically, the maximum difference in x and y coordinates between the left ankle joint and the left hip joint is taken, and a preset boundary margin is added to form a rectangular box. The boundary margin is dynamically set according to the differences in user body shape and movement range in the actual application scenario, and is usually 10% to 15% of the key point spacing. For example, in this application, the boundary margin can be set to 12% of the key point spacing to balance tracking robustness and region compactness.

[0046] Among them, the OpenPose algorithm is an existing real-time multi-person pose estimation algorithm based on convolutional neural networks. It can simultaneously detect multiple key joints of the human body and their connection relationships from a single RGB image. In this embodiment, the OpenPose algorithm is used to accurately extract the coordinates of the key points of the target user's lower limbs from the first frame image. Specifically, it performs feature extraction and key point confidence map regression on the input image through a multi-stage convolutional neural network, and outputs the two-dimensional coordinate positions of multiple key skeletal points including the left hip joint, right hip joint, left knee joint, right knee joint, left ankle joint, and right ankle joint.

[0047] The real-time image data obtained in step S10 provides high-quality, continuous, and lower limb-focused visual input for subsequent gait coordination feature extraction, and step S20 is then performed based on this.

[0048] S20: Extract the gait coordination features of the target user from the real-time image data, wherein the gait coordination features include at least the gait phase difference of the two lower limbs and the gait period variability;

[0049] In this embodiment, the extraction of gait coordination features is based on the temporal trajectory of key joints of the lower limbs in multiple consecutive frames of images. Specifically, the skeletal key point sequence is reconstructed from the real-time image data obtained in the above steps to obtain the skeletal key point positions of the target user. Then, based on the skeletal key point positions and multiple consecutive gait cycles, the gait phase difference and gait cycle variability of the lower limbs are calculated to obtain the gait coordination features of the target user.

[0050] Specifically, step S20 in the method includes:

[0051] Human skeletal key point detection is performed on each frame of the real-time image data to obtain the skeletal key point positions of the target user. The skeletal key points include at least the left ankle joint, right ankle joint, left knee joint, right knee joint, left hip joint, and right hip joint.

[0052] Based on the location of the skeletal key points, the swing state of the left lower limb and the swing state of the right lower limb of the target user in each frame of the image are determined, and multiple consecutive gait cycles are determined based on the swing and support states of the left and right lower limbs. Each gait cycle includes one swing phase and one support phase of the same lower limb.

[0053] The phase difference of the gait of the two lower limbs is determined based on the swing state of the left lower limb and the swing state of the right lower limb.

[0054] The cycle duration of multiple consecutive gait cycles is obtained, the statistical dispersion of the cycle duration of the multiple consecutive gait cycles is calculated, and the statistical dispersion is used as the gait cycle variability of the target user.

[0055] The gait phase difference of the two lower limbs and the gait period variability are used as the gait coordination characteristics of the target user.

[0056] In the embodiments of this application, such as Figure 2 As shown, firstly, the same human skeleton key point detection algorithm as S10 is used to detect the skeleton key points of each frame of the real-time image data to obtain the skeleton key point positions of the target user. Similar to the above steps, the skeleton key points include at least the left ankle joint, right ankle joint, left knee joint, right knee joint, left hip joint, and right hip joint.

[0057] Secondly, based on the location of key skeletal points, the swing states of the left and right lower limbs of the target user in each frame of the image are determined. Based on one swing phase and one support phase of the same lower limb, a gait cycle is determined. Based on the swing states of the left and right lower limbs, the phase difference of the gait between the two lower limbs is determined. The phase difference of the gait between the two lower limbs represents the proportion of time misalignment when the two lower limbs enter the swing phase. Furthermore, the phase difference of the two lower limbs in the normal gait of a healthy adult is usually between 0.45 and 0.55. The larger the deviation, the worse the coordination of the two lower limbs.

[0058] Simultaneously, the cycle duration of multiple consecutive gait cycles is obtained, and the statistical dispersion is calculated. The statistical dispersion is used as the gait cycle variability of the target user. The statistical dispersion of the cycle duration can be measured by standard deviation or coefficient of variation.

[0059] For example, taking the coefficient of variation as an example, if the durations of five consecutive gait cycles are 1.02 seconds, 1.08 seconds, 0.95 seconds, 1.10 seconds, and 0.98 seconds, respectively, the coefficient of variation = standard deviation / mean ≈ 5.7%. The coefficient of variation of gait cycles is usually expressed as a percentage. The coefficient of variation of gait cycles in normal people is generally between 3% and 6%. A value exceeding 8% usually indicates a decline in gait stability.

[0060] Finally, the phase difference of the lower limb gait and the variability of the gait cycle are used as gait coordination features of the target user and are jointly input into the subsequent steady state recognition model.

[0061] Furthermore, based on the positions of the skeletal key points, the swing states of the target user's left and right lower limbs in each frame of the image are determined, and multiple consecutive gait cycles are determined based on the swing and support states of the left and right lower limbs, including:

[0062] Based on the positions of the left hip joint, left knee joint, and left ankle joint, determine whether the left lower limb is in the swing phase or the support phase in the current frame. Specifically, when the anterior-posterior displacement of the left ankle joint relative to the left hip joint in the sagittal plane exceeds a preset swing threshold, the left lower limb is determined to be in the swing phase; otherwise, the left lower limb is determined to be in the support phase.

[0063] Based on the positions of the right hip joint, right knee joint, and right ankle joint, determine whether the right lower limb is in the swing phase or the support phase in the current frame. Specifically, when the anterior-posterior displacement of the right ankle joint relative to the right hip joint in the sagittal plane exceeds a preset swing threshold, the right lower limb is determined to be in the swing phase; otherwise, the right lower limb is determined to be in the support phase.

[0064] One swing phase and one stance phase of the same lower limb are defined as one gait cycle, and multiple gait cycles are defined continuously.

[0065] In this embodiment, firstly, based on the positions of the left hip joint, left knee joint, and left ankle joint, the displacement of the left ankle joint relative to the left hip joint in the sagittal plane (i.e., the anterior-posterior direction of the human body) is calculated. When this displacement exceeds a preset swing threshold, the left lower limb is determined to be in the swing phase; otherwise, the left lower limb is determined to be in the support phase. The swing threshold is set based on the kinematic parameters of the normal human gait, typically 15% to 20% of the lower limb length. For example, when the average lower limb length is 90 cm, the swing threshold can be set to 13.5 to 18 cm. Similarly, based on the positions of the right hip joint, right knee joint, and right ankle joint, the position of the right lower limb in the swing or support phase is determined.

[0066] Specifically, the swing phase refers to the stage where the lower limb on that side leaves the ground and steps forward, while the stance phase refers to the stage where the lower limb on that side contacts the ground and supports the body weight. The alternation of these two phases constitutes the basic rhythm of gait.

[0067] Secondly, based on the swing and support states of the left and right lower limbs, multiple consecutive gait cycles are determined. A gait cycle is defined as the complete process of the same lower limb from the start of one swing phase to the start of the next swing phase, including one swing phase and one support phase. For example, taking the left lower limb as a reference, the time from when the left lower limb enters the swing phase to when it enters the next swing phase is a complete left lower limb gait cycle. Usually, 5 to 10 consecutive gait cycles are taken. To balance the statistical reliability and real-time response speed of gait analysis, this embodiment takes 5 consecutive gait cycles for feature calculation. 5 cycles are sufficient to calculate an effective estimate of the gait cycle variation coefficient, while avoiding feedback delay caused by too many cycles.

[0068] The determination of the gait phase difference between the two lower limbs based on the swing states of the left and right lower limbs includes:

[0069] Obtain the first moment when the left lower limb enters the swing phase and the second moment when the right lower limb enters the swing phase within each gait cycle;

[0070] Calculate the time difference between the second time point and the first time point to obtain the bi-lower limb gait phase time difference corresponding to each gait cycle;

[0071] Divide the bi-leg gait phase time difference corresponding to each gait cycle by the cycle duration of the corresponding gait cycle to obtain the bi-leg gait phase difference corresponding to each gait cycle.

[0072] Calculate the average value of the phase difference of the lower limb gait corresponding to multiple gait cycles, and use the average value as the phase difference of the lower limb gait.

[0073] In this embodiment, firstly, the first moment when the left lower limb enters the swing phase and the second moment when the right lower limb enters the swing phase within each gait cycle are obtained. Specifically, the starting point of the swing phase is determined by detecting the peak value of the sagittal displacement change rate of the left and right ankle joints relative to their respective hip joints in consecutive frames. When the displacement change rate changes from negative to positive and exceeds a preset dynamic threshold, it is determined to be the start time of the swing phase. The preset dynamic threshold is dynamically adjusted according to the individual gait speed of the user and is usually set to 1.5 times the standard deviation of the mean displacement change rate.

[0074] Secondly, calculate the time difference between the second moment and the first moment to obtain the phase time difference of the lower limb gait corresponding to the gait cycle. For example, in a certain cycle, the left lower limb enters the swing phase at t=0.32 seconds and the right lower limb enters the swing phase at t=0.81 seconds, then the phase time difference is 0.49 seconds.

[0075] Next, the phase time difference is divided by the total duration of the corresponding gait cycle to obtain the normalized lower limb gait phase difference. This normalization process eliminates the influence of individual walking speed differences on phase assessment, making the phase difference comparable across users.

[0076] Finally, the arithmetic mean of the phase differences calculated from multiple consecutive effective gait cycles is taken as the final lower limb gait phase difference feature value used for state determination, providing input for the subsequent stable state recognition model.

[0077] S30: Input the gait coordination features into a pre-trained lightweight stable state recognition model, and output the center of gravity stability level of the target user from the lightweight stable state recognition model;

[0078] In this embodiment, the lightweight stable state recognition model includes a first-level classifier, a second-level classifier, and a gating module. Its input features are the phase difference of the lower limb gait and the coefficient of variation of the gait cycle. The output is the center of gravity stability level of the target user, specifically including four levels: stable, slightly unstable, moderately unstable, and highly unstable. It should be noted that before inputting into the lightweight stable state recognition model, the gait coordination features of multiple consecutive sampling times are subjected to moving average filtering to eliminate single-frame detection noise. Then, the gait cycle variability is normalized to the [0,1] interval. The normalization method can be maximum-minimum normalization.

[0079] The pre-training process of the lightweight stable state recognition model includes:

[0080] Multiple historical gait samples are acquired. Each historical gait sample includes historical gait coordination features and a corresponding center of gravity stability level label. The historical gait coordination features include historical lower limb gait phase difference and historical gait period variability. The center of gravity stability level label includes stable, mild instability, moderate instability and severe instability.

[0081] A lightweight steady-state recognition model is constructed, which includes a first-level classifier, a second-level classifier, and a gating module.

[0082] Historical gait samples labeled as stable at the center of gravity stability level are used as positive samples for the first-level classifier, and historical gait samples labeled as unstable at the center of gravity stability level are used as negative samples for the first-level classifier. The first-level classifier is trained in a supervised manner until it is verified to converge, and the trained first-level classifier is obtained. Here, the unstable refers to mild instability, moderate instability or high instability.

[0083] Historical gait samples labeled as mild instability, moderate instability, and severe instability are used as training samples for the second-level classifier. The second-level classifier is trained in a supervised manner until it is verified to converge, and the trained second-level classifier is obtained.

[0084] The trained first-level classifier, the trained second-level classifier, and the gating module are combined to obtain a pre-trained lightweight stable state recognition model.

[0085] In this embodiment of the application, this step uses a pre-trained lightweight stable state recognition model to classify gait coordination features and output the current center of gravity stability level of the target user. The pre-training process of the lightweight stable state recognition model is as follows:

[0086] First, multiple historical gait samples are obtained. Each historical gait sample includes historical gait coordination features and corresponding center of gravity stability level labels. Historical gait coordination features include historical lower limb gait phase difference and historical gait period variability. Center of gravity stability level labels are marked according to clinical assessment criteria and are divided into stable, mild instability, moderate instability and high instability. The number of samples is usually no less than 2,000 to ensure the generalization ability of the model.

[0087] Secondly, a lightweight stable state recognition model is constructed. This model adopts a two-level cascaded architecture, including a first-level classifier, a second-level classifier, and a gating module.

[0088] Furthermore, the training process is divided into two stages. In the first stage, historical gait samples labeled as stable at the center of gravity are used as positive samples for the first-level classifier, while historical gait samples labeled as unstable at the center of gravity (i.e., slightly unstable, moderately unstable, or highly unstable) are used as negative samples for the first-level classifier. The first-level classifier is trained in a supervised manner until the validation set converges. For example, the training uses the cross-entropy loss function, the optimizer is Adam, the learning rate is set to 0.001, the number of training rounds is set to 100, and convergence is considered when the validation set accuracy reaches 95% or higher.

[0089] In the second stage, historical gait samples labeled with the stability level of the center of gravity as mild instability, moderate instability, and severe instability are used as training samples for the second-level classifier. The second-level classifier is trained in a supervised manner until the validation set converges. Using the same training configuration, convergence can be considered when the multi-class accuracy on the validation set reaches more than 90%.

[0090] Finally, the trained first-level classifier, the trained second-level classifier, and the gating module are combined to obtain a pre-trained lightweight stable state recognition model.

[0091] Furthermore, a lightweight stable state identification model is constructed, including:

[0092] The first-level classifier is a lightweight binary classification model, including a first input layer, a first hidden layer and a first output layer. The first input layer receives gait coordination features, the first hidden layer performs a nonlinear transformation on the gait coordination features, and the first output layer outputs the probabilities of stable and unstable states, and takes the state with the higher probability as the first-level classification result.

[0093] The second-level classifier is a lightweight multi-classification model, including a second input layer, a second hidden layer, and a second output layer. The second input layer receives the gait coordination features, the second hidden layer performs a nonlinear transformation on the gait coordination features, and the second output layer outputs the probabilities of mild instability, moderate instability, and high instability, and takes the state with the highest probability as the second-level classification result.

[0094] The gating module is used to receive the first-level classification result, determine whether to activate the second-level classifier based on the first-level classification result, and output the centroid stability level based on the first-level classification result and / or the second-level classification result.

[0095] Wherein, the number of hidden layer nodes in the first-level classifier is less than the number of hidden layer nodes in the second-level classifier, so that the computational cost of the first-level classifier is less than that of the second-level classifier;

[0096] The decision to activate the second-level classifier based on the first-level classification result includes: when the first-level classification result is in a stable state, the gating module does not activate the second-level classifier and uses the stable state as the centroid stability level; when the first-level classification result is in an unstable state, the gating module activates the second-level classifier and uses the second-level classification result as the centroid stability level.

[0097] In this embodiment, the first-level classifier is a lightweight binary classification model. Its network structure includes a first input layer, a first hidden layer, and a first output layer. The first input layer receives gait coordination features, which are two-dimensional vectors, namely the phase difference of the lower limbs and the variability of the gait cycle.

[0098] For example, the first hidden layer performs a nonlinear transformation on the gait coordination features. The number of hidden layer nodes is set to 8, and the ReLU activation function is used. The first output layer uses the Softmax activation function to output the probabilities of stable and unstable states, and the state with the higher probability is used as the first-level classification result. For example, the first-level classifier has about 30 parameters, and the computational cost is extremely small, making it suitable for high-frequency use.

[0099] The second-level classifier is a lightweight multi-classification model. Its network structure includes a second input layer, a second hidden layer, and a second output. The second input layer also receives gait coordination features, and the second hidden layer performs a non-linear transformation on the gait coordination features. The number of hidden layer nodes is set to 32, and the ReLU activation function is used. The second output layer uses the Softmax activation function to output the probabilities of three categories: mild instability, moderate instability, and high instability. The state with the highest probability is taken as the second-level classification result. For example, the second-level classifier has approximately 130 parameters, slightly more than the first-level classifier, but it is only activated when needed.

[0100] In this design, the number of hidden layer nodes in the first-level classifier is less than that in the second-level classifier, making the computational cost of the first-level classifier less than that of the second-level classifier. This design realizes a hierarchical processing strategy of first coarse screening and then fine judgment.

[0101] The gating module receives the first-level classification result and determines whether to activate the second-level classifier based on the first-level classification result. The specific decision logic is as follows: when the first-level classification result is a stable state, the gating module does not activate the second-level classifier and directly outputs the stable state as the centroid stability level; when the first-level classification result is an unstable state, the gating module activates the second-level classifier and outputs the classification result of the second-level classifier, i.e., mild instability, moderate instability, or high instability, as the centroid stability level.

[0102] Based on step S30, the current center of gravity stability level of the target user can be obtained. This level serves as the core basis for caregiver action decisions and directly drives the generation and execution of subsequent caregiver instructions.

[0103] S40: Based on the center of gravity stability level, match a target care instruction from a preset care instruction set. The target care instruction is used to adjust at least one of the following: the movement speed, movement direction, or handrail height of the health care robot.

[0104] In this embodiment, a target care instruction is matched from a preset care instruction set according to the center of gravity stability level. The preset care instruction set is divided into four levels according to the center of gravity stability level, corresponding to stable, mild instability, moderate instability and high instability states respectively. Each level contains a set of parameterized action adjustment strategies, which are used to dynamically adjust the behavior of the rehabilitation robot to adapt to the user's current gait risk level.

[0105] Specifically, step S40 in the method includes:

[0106] Based on the center of gravity stability level, a target care instruction is matched from a preset care instruction set. The target care instruction is used to adjust at least one of the following: the movement speed, movement direction, or handrail height of the health care robot.

[0107] The preset caregiver instruction set is constructed in the following way:

[0108] The system acquires the command execution capability of the health care robot, wherein the command execution capability includes at least one of the following: movement speed adjustment command, movement direction adjustment command, and handrail height adjustment command that the health care robot can execute;

[0109] Based on the center of gravity stability level, different combinations of target care instructions are assigned to different center of gravity stability levels;

[0110] Store the correspondence between the center of gravity stability level and the target care instruction combination as a preset care instruction set.

[0111] In this embodiment of the application, the step matches the corresponding target care instruction from the preset care instruction set according to the center of gravity stability level output in step S30, wherein the preset care instruction set is constructed in the following way:

[0112] First, the command execution capability of the health and wellness robot needs to be obtained, including at least one of the following commands: movement speed adjustment command, movement direction adjustment command, and handrail height adjustment command. Different models of robots have different execution capabilities, which need to be determined based on the specific hardware configuration.

[0113] Secondly, based on the center of gravity stability level, different combinations of target care instructions are assigned to different center of gravity stability levels, as shown in Table 1, Instruction Assignment Strategy Table:

[0114] Center of gravity stability level Movement speed adjustment Adjusting the direction of movement Armrest height adjustment Stablize Maintain current speed Maintain current direction Maintain current altitude Mild instability Reduce speed by 10% to 20%. Maintain current direction Raise the handrail height by 5-10 centimeters Moderate instability Reduce speed by 30% to 50% Deflect 5°~10° in the opposite direction to the user's center of gravity offset. Raise the handrail height by 10-15 cm High instability Stop moving Maintain current direction Raise the handrail to its highest position.

[0115] Table 1 Instruction Allocation Strategy Table

[0116] As shown in Table 1, the walking speed adjustment values ​​are set according to the "Reference Standard for Gait Speed ​​of the Elderly". The natural walking speed of healthy elderly people is about 0.8~1.2m / s; the speed should be reduced to 0.4~0.7m / s in an unstable state. The handrail height adjustment is based on ergonomics. Every 5 cm increase in handrail height can lower the center of gravity by about 2 cm, thereby effectively increasing the stability of the support surface. However, an increase of more than 15 cm will cause excessive flexion of the user's elbow joint, so the upper limit is set at 15 cm.

[0117] Finally, the correspondence between the center of gravity stability level and the target care instruction combination is stored as a preset care instruction set. For example, the target care instruction for mild instability is to reduce the movement speed by 15% and increase the handrail height by 8 cm; the target care instruction for moderate instability is to reduce the movement speed by 40%, turn the steering wheel by 5 degrees, and increase the handrail height by 12 cm.

[0118] S50: The target care instruction is sent to the health care robot for execution, and after execution, the real-time image data is acquired to update the gait coordination characteristics.

[0119] In this embodiment of the application, the target care instruction obtained by matching is sent to the execution mechanism through the robot operating system or proprietary communication protocol, including the mobile chassis, steering mechanism and handrail lifting mechanism, etc.

[0120] After the instruction is executed, the health care robot returns to step S10 and continues to collect real-time image data of the target user through the tracking vision sensor. It then re-extracts the gait coordination features and re-inputs them into the stable state recognition model to evaluate the changes in the center of gravity stability level. This forms a closed-loop feedback control, enabling the robot's companion actions to be dynamically optimized according to the real-time changes in the user's gait state.

[0121] In summary, compared with existing technologies, this application effectively solves the problem of delayed caregiving response caused by discontinuous perception and single evaluation dimension in traditional health care robots by introducing a tracking visual sensor to collect data on the user's lower limb area and combining it with a gait coordination feature extraction method based on the temporal trajectory of skeletal key points.

[0122] In summary, the embodiments of this application have at least the following technical effects:

[0123] This application provides an AI-assisted intelligent optimization method for the companion movements of a rehabilitation robot. First, it utilizes a tracking visual sensor mounted on the rehabilitation robot to achieve continuous, unobstructed tracking of the target user's lower limbs, avoiding the burden of wearing wearable devices and the challenges to user compliance. Second, it extracts the gait phase difference and gait period variability of both lower limbs from real-time image data to construct multi-dimensional gait coordination features, improving the early detection sensitivity of instability. Third, it employs a two-level lightweight classification architecture combined with a gating mechanism for a stable state recognition model, ensuring high discrimination accuracy while maintaining computational efficiency. Finally, it dynamically matches and executes appropriate companion commands based on the stability level of the center of gravity, enabling the robot to actively adjust its movement speed, direction, or handrail height.

[0124] Through the above technical solution, this application realizes a closed-loop optimization mechanism from gait perception and status recognition to caregiving response, achieving coordinated response with the user's gait behavior, thereby effectively reducing the risk of falls and improving the intelligence, personalization and safety of caregiving services.

[0125] Example 2, as Figure 3 As shown, based on the same inventive concept as the AI ​​vision-assisted intelligent optimization method for the companion movements of a healthcare robot provided in Embodiment 1, this application also provides an AI vision-assisted intelligent optimization system for the companion movements of a healthcare robot, including:

[0126] Image data acquisition module 11 is used to acquire real-time image data of the target user through a tracking vision sensor installed on the health and wellness robot;

[0127] The gait coordination judgment module 12 is used to extract the gait coordination features of the target user from the real-time image data. The gait coordination features include at least the gait phase difference of the two lower limbs and the gait period variability.

[0128] The recognition model training module 13 is used to input the gait coordination features into a pre-trained lightweight stable state recognition model, and the lightweight stable state recognition model outputs the center of gravity stability level of the target user.

[0129] The care instruction matching module 14 is used to match a target care instruction from a preset care instruction set according to the center of gravity stability level. The target care instruction is used to adjust at least one of the following: the moving speed, the moving direction, or the handrail height of the health care robot.

[0130] The care instruction execution module 15 is used to send the target care instruction to the health care robot for execution, and continue to acquire the real-time image data after execution to update the gait coordination features.

[0131] In one embodiment, the image data acquisition module 11 is specifically used for:

[0132] The tracking vision sensor includes an image acquisition unit and an attitude adjustment unit, wherein the image acquisition unit is used to acquire images, and the attitude adjustment unit is used to adjust the pitch angle, yaw angle, and focal length of the image acquisition unit.

[0133] The tracking vision sensor detects the target user and identifies the target user's lower limb area;

[0134] Based on the position and size of the lower limb region in the image, the pitch angle, yaw angle and focal length of the image acquisition unit are adjusted by the posture adjustment unit to keep the lower limb region within the preset area of ​​the image acquisition screen, and to keep the lower limb region intact and clear in the image;

[0135] During the movement of the target user, the tracking vision sensor continuously acquires real-time image data containing the lower limb region, and the real-time image data includes multiple frames of images arranged in chronological order.

[0136] The tracking visual sensor detects the target user and identifies the target user's lower limb region, including:

[0137] Human skeletal key point detection is performed on the first frame image acquired by the image acquisition unit to obtain the skeletal key point positions of the target user. The skeletal key point positions include at least the positions of the left hip joint, right hip joint, left knee joint, right knee joint, left ankle joint, and right ankle joint.

[0138] Based on the location of the key skeletal points, a bounding box is determined that surrounds the left and right lower limbs, and the bounding box is used as the lower limb region;

[0139] In subsequent frames, a lightweight target tracking algorithm is used to track the lower limb region, and the position and size of the lower limb region are updated according to the tracked position.

[0140] When the tracking confidence of the lightweight target tracking algorithm is lower than the preset confidence threshold, the human skeleton key point detection is re-performed. After the location of the skeleton key point is re-acquired and the lower limb region is re-determined, the tracking of the lower limb region is resumed.

[0141] In one embodiment, the gait coordination judgment module 12 is specifically used for:

[0142] Human skeletal key point detection is performed on each frame of the real-time image data to obtain the skeletal key point positions of the target user. The skeletal key points include at least the left ankle joint, right ankle joint, left knee joint, right knee joint, left hip joint, and right hip joint.

[0143] Based on the location of the skeletal key points, the swing state of the left lower limb and the swing state of the right lower limb of the target user in each frame of the image are determined, and multiple consecutive gait cycles are determined based on the swing and support states of the left and right lower limbs. Each gait cycle includes one swing phase and one support phase of the same lower limb.

[0144] The phase difference of the gait of the two lower limbs is determined based on the swing state of the left lower limb and the swing state of the right lower limb.

[0145] The cycle duration of multiple consecutive gait cycles is obtained, the statistical dispersion of the cycle duration of the multiple consecutive gait cycles is calculated, and the statistical dispersion is used as the gait cycle variability of the target user.

[0146] The gait phase difference of the two lower limbs and the gait period variability are used as the gait coordination characteristics of the target user.

[0147] Furthermore, based on the positions of the skeletal key points, the swing states of the target user's left and right lower limbs in each frame of the image are determined, and multiple consecutive gait cycles are determined based on the swing and support states of the left and right lower limbs, including:

[0148] Based on the positions of the left hip joint, left knee joint, and left ankle joint, determine whether the left lower limb is in the swing phase or the support phase in the current frame. Specifically, when the anterior-posterior displacement of the left ankle joint relative to the left hip joint in the sagittal plane exceeds a preset swing threshold, the left lower limb is determined to be in the swing phase; otherwise, the left lower limb is determined to be in the support phase.

[0149] Based on the positions of the right hip joint, right knee joint, and right ankle joint, determine whether the right lower limb is in the swing phase or the support phase in the current frame. Specifically, when the anterior-posterior displacement of the right ankle joint relative to the right hip joint in the sagittal plane exceeds a preset swing threshold, the right lower limb is determined to be in the swing phase; otherwise, the right lower limb is determined to be in the support phase.

[0150] One swing phase and one stance phase of the same lower limb are defined as one gait cycle, and multiple gait cycles are defined continuously.

[0151] Further, based on the swing states of the left and right lower limbs, the phase difference of the gait between the two lower limbs is determined, including:

[0152] Obtain the first moment when the left lower limb enters the swing phase and the second moment when the right lower limb enters the swing phase within each gait cycle;

[0153] Calculate the time difference between the second time point and the first time point to obtain the bi-lower limb gait phase time difference corresponding to each gait cycle;

[0154] Divide the bi-leg gait phase time difference corresponding to each gait cycle by the cycle duration of the corresponding gait cycle to obtain the bi-leg gait phase difference corresponding to each gait cycle.

[0155] Calculate the average value of the phase difference of the lower limb gait corresponding to multiple gait cycles, and use the average value as the phase difference of the lower limb gait.

[0156] Furthermore, in one embodiment of the application, the pre-training process of the lightweight steady-state recognition model includes:

[0157] Multiple historical gait samples are acquired. Each historical gait sample includes historical gait coordination features and a corresponding center of gravity stability level label. The historical gait coordination features include historical lower limb gait phase difference and historical gait period variability. The center of gravity stability level label includes stable, mild instability, moderate instability and severe instability.

[0158] A lightweight steady-state recognition model is constructed, which includes a first-level classifier, a second-level classifier, and a gating module.

[0159] Historical gait samples labeled as stable at the center of gravity stability level are used as positive samples for the first-level classifier, and historical gait samples labeled as unstable at the center of gravity stability level are used as negative samples for the first-level classifier. The first-level classifier is trained in a supervised manner until it is verified to converge, and the trained first-level classifier is obtained. Here, the unstable refers to mild instability, moderate instability or high instability.

[0160] Historical gait samples labeled as mild instability, moderate instability, and severe instability are used as training samples for the second-level classifier. The second-level classifier is trained in a supervised manner until it is verified to converge, and the trained second-level classifier is obtained.

[0161] The trained first-level classifier, the trained second-level classifier, and the gating module are combined to obtain a pre-trained lightweight stable state recognition model.

[0162] The construction of a lightweight stable state recognition model includes:

[0163] The first-level classifier is a lightweight binary classification model, including a first input layer, a first hidden layer and a first output layer. The first input layer receives gait coordination features, the first hidden layer performs a nonlinear transformation on the gait coordination features, and the first output layer outputs the probabilities of stable and unstable states, and takes the state with the higher probability as the first-level classification result.

[0164] The second-level classifier is a lightweight multi-classification model, including a second input layer, a second hidden layer, and a second output layer. The second input layer receives the gait coordination features, the second hidden layer performs a nonlinear transformation on the gait coordination features, and the second output layer outputs the probabilities of mild instability, moderate instability, and high instability, and takes the state with the highest probability as the second-level classification result.

[0165] The gating module is used to receive the first-level classification result, determine whether to activate the second-level classifier based on the first-level classification result, and output the centroid stability level based on the first-level classification result and / or the second-level classification result.

[0166] Wherein, the number of hidden layer nodes in the first-level classifier is less than the number of hidden layer nodes in the second-level classifier, so that the computational cost of the first-level classifier is less than that of the second-level classifier;

[0167] The decision to activate the second-level classifier based on the first-level classification result includes: when the first-level classification result is in a stable state, the gating module does not activate the second-level classifier and uses the stable state as the centroid stability level; when the first-level classification result is in an unstable state, the gating module activates the second-level classifier and uses the second-level classification result as the centroid stability level.

[0168] In one embodiment of the application, the caregiver instruction matching module 14 is specifically used for:

[0169] Based on the center of gravity stability level, a target care instruction is matched from a preset care instruction set. The target care instruction is used to adjust at least one of the following: the movement speed, movement direction, or handrail height of the health care robot.

[0170] The preset caregiver instruction set is constructed in the following way:

[0171] The system acquires the command execution capability of the health care robot, wherein the command execution capability includes at least one of the following: movement speed adjustment command, movement direction adjustment command, and handrail height adjustment command that the health care robot can execute;

[0172] Based on the center of gravity stability level, different combinations of target care instructions are assigned to different center of gravity stability levels;

[0173] Store the correspondence between the center of gravity stability level and the target care instruction combination as a preset care instruction set.

Claims

1. A method for intelligently optimizing the movements of AI vision-assisted elderly care robot companions, characterized in that, The method includes: Real-time image data of the target user is collected by a tracking visual sensor installed on the health and wellness robot; Gait coordination features of the target user are extracted from the real-time image data, and the gait coordination features include at least the gait phase difference of the two lower limbs and the gait period variability; The gait coordination features are input into a pre-trained lightweight stable state recognition model, which then outputs the center of gravity stability level of the target user. Based on the center of gravity stability level, a target care instruction is matched from a preset care instruction set. The target care instruction is used to adjust at least one of the following: the movement speed, movement direction, or handrail height of the health care robot. The target care instruction is sent to the health care robot for execution, and after execution, the robot continues to acquire the real-time image data to update the gait coordination characteristics.

2. The AI ​​vision-assisted intelligent optimization method for the companionship movements of a healthcare robot as described in claim 1, characterized in that, Real-time image data of the target user is collected by tracking vision sensors installed on the health and wellness robot, including: The tracking vision sensor includes an image acquisition unit and an attitude adjustment unit, wherein the image acquisition unit is used to acquire images, and the attitude adjustment unit is used to adjust the pitch angle, yaw angle, and focal length of the image acquisition unit. The tracking vision sensor detects the target user and identifies the target user's lower limb area; Based on the position and size of the lower limb region in the image, the pitch angle, yaw angle and focal length of the image acquisition unit are adjusted by the posture adjustment unit to keep the lower limb region within the preset area of ​​the image acquisition screen, and to keep the lower limb region intact and clear in the image; During the movement of the target user, the tracking vision sensor continuously acquires real-time image data containing the lower limb region, and the real-time image data includes multiple frames of images arranged in chronological order.

3. The AI ​​vision-assisted intelligent optimization method for the companionship movements of a healthcare robot as described in claim 2, characterized in that, The tracking vision sensor detects the target user and identifies the target user's lower limb region, including: Human skeletal key point detection is performed on the first frame image acquired by the image acquisition unit to obtain the skeletal key point positions of the target user. The skeletal key point positions include at least the positions of the left hip joint, right hip joint, left knee joint, right knee joint, left ankle joint, and right ankle joint. Based on the location of the key skeletal points, a bounding box is determined that surrounds the left and right lower limbs, and the bounding box is used as the lower limb region; In subsequent frames, a lightweight target tracking algorithm is used to track the lower limb region, and the position and size of the lower limb region are updated according to the tracked position. When the tracking confidence of the lightweight target tracking algorithm is lower than the preset confidence threshold, the human skeleton key point detection is re-performed. After the location of the skeleton key point is re-acquired and the lower limb region is re-determined, the tracking of the lower limb region is resumed.

4. The method for intelligent optimization of the movements of a healthcare robot with AI vision assistance according to claim 1, characterized in that, Extracting the gait coordination features of the target user from the real-time image data includes: Human skeletal key point detection is performed on each frame of the real-time image data to obtain the skeletal key point positions of the target user. The skeletal key points include at least the left ankle joint, right ankle joint, left knee joint, right knee joint, left hip joint, and right hip joint. Based on the location of the skeletal key points, the swing state of the left lower limb and the swing state of the right lower limb of the target user in each frame of the image are determined, and multiple consecutive gait cycles are determined based on the swing and support states of the left and right lower limbs. Each gait cycle includes one swing phase and one support phase of the same lower limb. The phase difference of the gait of the two lower limbs is determined based on the swing state of the left lower limb and the swing state of the right lower limb. The cycle duration of multiple consecutive gait cycles is obtained, the statistical dispersion of the cycle duration of the multiple consecutive gait cycles is calculated, and the statistical dispersion is used as the gait cycle variability of the target user. The gait phase difference of the two lower limbs and the gait period variability are used as the gait coordination characteristics of the target user.

5. The AI ​​vision-assisted intelligent optimization method for the companionship movements of a health and wellness robot according to claim 4, characterized in that, Based on the locations of the skeletal key points, the swing states of the target user's left and right lower limbs in each frame of the image are determined. Based on the swing and support states of the left and right lower limbs, multiple consecutive gait cycles are determined, including: Based on the positions of the left hip joint, left knee joint, and left ankle joint, determine whether the left lower limb is in the swing phase or the support phase in the current frame. Specifically, when the anterior-posterior displacement of the left ankle joint relative to the left hip joint in the sagittal plane exceeds a preset swing threshold, the left lower limb is determined to be in the swing phase; otherwise, the left lower limb is determined to be in the support phase. Based on the positions of the right hip joint, right knee joint, and right ankle joint, determine whether the right lower limb is in the swing phase or the support phase in the current frame. Specifically, when the anterior-posterior displacement of the right ankle joint relative to the right hip joint in the sagittal plane exceeds a preset swing threshold, the right lower limb is determined to be in the swing phase; otherwise, the right lower limb is determined to be in the support phase. One swing phase and one stance phase of the same lower limb are defined as one gait cycle, and multiple gait cycles are defined continuously.

6. The AI ​​vision-assisted intelligent optimization method for the companionship movements of a health and wellness robot according to claim 4, characterized in that, Determining the gait phase difference between the two lower limbs based on the swing states of the left and right lower limbs includes: Obtain the first moment when the left lower limb enters the swing phase and the second moment when the right lower limb enters the swing phase within each gait cycle; Calculate the time difference between the second time point and the first time point to obtain the bi-lower limb gait phase time difference corresponding to each gait cycle; Divide the bi-leg gait phase time difference corresponding to each gait cycle by the cycle duration of the corresponding gait cycle to obtain the bi-leg gait phase difference corresponding to each gait cycle. Calculate the average value of the phase difference of the lower limb gait corresponding to multiple gait cycles, and use the average value as the phase difference of the lower limb gait.

7. The AI ​​vision-assisted intelligent optimization method for the companionship movements of a health and wellness robot according to claim 1, characterized in that, The pre-training process of the lightweight steady-state recognition model includes: Multiple historical gait samples are acquired. Each historical gait sample includes historical gait coordination features and a corresponding center of gravity stability level label. The historical gait coordination features include historical lower limb gait phase difference and historical gait period variability. The center of gravity stability level label includes stable, mild instability, moderate instability and severe instability. A lightweight steady-state recognition model is constructed, which includes a first-level classifier, a second-level classifier, and a gating module. Historical gait samples labeled as stable at the center of gravity stability level are used as positive samples for the first-level classifier, and historical gait samples labeled as unstable at the center of gravity stability level are used as negative samples for the first-level classifier. The first-level classifier is trained in a supervised manner until it is verified to converge, and the trained first-level classifier is obtained. Here, the unstable refers to mild instability, moderate instability or high instability. Historical gait samples labeled as mild instability, moderate instability, and severe instability are used as training samples for the second-level classifier. The second-level classifier is trained in a supervised manner until it is verified to converge, and the trained second-level classifier is obtained. The trained first-level classifier, the trained second-level classifier, and the gating module are combined to obtain a pre-trained lightweight stable state recognition model.

8. The method for intelligent optimization of the movements of a healthcare robot with AI vision assistance according to claim 7, characterized in that, Construct a lightweight steady-state identification model, including: The first-level classifier is a lightweight binary classification model, including a first input layer, a first hidden layer and a first output layer. The first input layer receives gait coordination features, the first hidden layer performs a nonlinear transformation on the gait coordination features, and the first output layer outputs the probabilities of stable and unstable states, and takes the state with the higher probability as the first-level classification result. The second-level classifier is a lightweight multi-classification model, including a second input layer, a second hidden layer, and a second output layer. The second input layer receives the gait coordination features, the second hidden layer performs a nonlinear transformation on the gait coordination features, and the second output layer outputs the probabilities of mild instability, moderate instability, and high instability, and takes the state with the highest probability as the second-level classification result. The gating module is used to receive the first-level classification result, determine whether to activate the second-level classifier based on the first-level classification result, and output the centroid stability level based on the first-level classification result and / or the second-level classification result. Wherein, the number of hidden layer nodes in the first-level classifier is less than the number of hidden layer nodes in the second-level classifier, so that the computational cost of the first-level classifier is less than that of the second-level classifier; The decision to activate the second-level classifier based on the first-level classification result includes: when the first-level classification result is in a stable state, the gating module does not activate the second-level classifier and uses the stable state as the centroid stability level; when the first-level classification result is in an unstable state, the gating module activates the second-level classifier and uses the second-level classification result as the centroid stability level.

9. The method for intelligent optimization of the movements of a healthcare robot with AI vision assistance according to claim 1, characterized in that, Based on the aforementioned center of gravity stability level, target care instructions are matched from a preset care instruction set, including: Based on the center of gravity stability level, a target care instruction is matched from a preset care instruction set. The target care instruction is used to adjust at least one of the following: the movement speed, movement direction, or handrail height of the health care robot. The preset caregiver instruction set is constructed in the following way: The system acquires the command execution capability of the health care robot, wherein the command execution capability includes at least one of the following: movement speed adjustment command, movement direction adjustment command, and handrail height adjustment command that the health care robot can execute; Based on the center of gravity stability level, different combinations of target care instructions are assigned to different center of gravity stability levels; Store the correspondence between the center of gravity stability level and the target care instruction combination as a preset care instruction set.

10. An AI-assisted vision-based intelligent optimization system for the movements of a health and wellness robot companion, characterized in that: The method for intelligently optimizing the movements of a healthcare robot with AI visual assistance as described in any one of claims 1 to 9 includes: The image data acquisition module is used to acquire real-time image data of the target user through a tracking vision sensor installed on the health and wellness robot. A gait coordination judgment module is used to extract the gait coordination features of the target user from the real-time image data. The gait coordination features include at least the gait phase difference of the two lower limbs and the gait period variability. The recognition model training module is used to input the gait coordination features into a pre-trained lightweight stable state recognition model, and the lightweight stable state recognition model outputs the center of gravity stability level of the target user. The care instruction matching module is used to match a target care instruction from a preset care instruction set according to the center of gravity stability level. The target care instruction is used to adjust at least one of the following: the moving speed, the moving direction, or the handrail height of the health care robot. The care instruction execution module is used to send the target care instruction to the health care robot for execution, and continue to acquire the real-time image data after execution to update the gait coordination features.

Images