Non-inductive calibration and pose screening-based home-based elderly care non-inductive monitoring method and system

Abstract

The application relates to the technical field of intelligent old-age care health monitoring, in particular to a home old-age care non-inductive monitoring method and system based on non-inductive calibration and pose screening, which comprises the following steps: acquiring multi-modal original data of a monitoring environment and a monitoring object; clustering the multi-modal original data by using a clustering algorithm, and non-inductively calibrating the monitoring object according to a clustering cluster center and a corresponding text description label; based on an observation effectiveness and a hidden non-inductive double-constraint model, combining an environmental semantic map, a real-time pose of the monitoring object and a visual field range, and calculating an optimal observation pose of a robot; constructing a multi-layer cost map containing a semantic hidden cost layer, and executing navigation control for avoiding the visual field of the monitoring object according to the optimal observation pose; the semantic hidden cost layer sets the visual field area of the monitoring object as a high-cost area, and sets an environmental shelter shelter area as a gain area. The application can significantly reduce the interference and visual disturbance to the daily life of the old people.

Description

Technical Field

[0001] This application relates to the field of intelligent elderly care health monitoring technology, specifically to a method and system for non-intrusive home-based elderly care monitoring based on non-intrusive calibration and posture screening. Background Technology

[0002] With the arrival of an aging society, home-based elderly care monitoring robots have become an important tool to protect the lives and safety of the elderly. Existing monitoring robot technology mainly uses cameras and various sensors to achieve real-time monitoring of the elderly's activity status.

[0003] However, existing home monitoring technologies have the following significant drawbacks in practical applications: Traditional monitoring systems typically require high-resolution raw image streams of faces or bodies for identification. Once this sensitive data is uploaded to the cloud or illegally intercepted, it poses a serious threat to the privacy of the elderly. Existing robot navigation algorithms primarily focus on path minimization or obstacle avoidance safety, neglecting the psychological resilience of the elderly. If a robot frequently starts, stops, or crosses the elderly person's field of vision during monitoring tasks, it can easily startle them, potentially leading to falls or other accidents. Traditional follow-up monitoring robots often remain within the elderly person's peripheral vision. This "shadow-like" monitoring method creates a strong sense of psychological pressure for the elderly, and prolonged exposure to such an environment can severely impact their mental health and living comfort. Summary of the Invention

[0004] The objective of this application is to provide a method for non-intrusive home-based elderly care monitoring based on non-intrusive calibration and pose selection, comprising: acquiring multimodal raw data of the monitoring environment and the monitored object; clustering the multimodal raw data using a clustering algorithm, and non-intrusively calibrating the monitored object according to the cluster centers and corresponding text description labels; calculating the optimal observation pose of the robot based on a dual-constraint model of observation effectiveness and stealth non-intrusiveness, combined with an environmental semantic map, the real-time pose of the monitored object, and the field of vision; constructing a multi-layer cost map including a semantic stealth cost layer, and executing navigation control to avoid the field of vision of the monitored object according to the optimal observation pose; wherein the semantic stealth cost layer sets the field of vision area of ​​the monitored object as a high-cost area and sets the environmental cover occlusion area as a gain area.

[0005] By adopting the above technical solution, the optimal observation pose is calculated by combining a dual-constraint model based on observation effectiveness and stealthless movement, and a semantic stealth cost layer is constructed to guide the robot to avoid the field of vision of the monitored object. This enables the robot to autonomously choose a hidden and inconspicuous observation position and movement path when performing monitoring tasks, which significantly reduces interference and visual disturbance to the daily life of the elderly.

[0006] Optionally, it also includes: extracting multimodal features from the original multimodal data and constructing a feature vector; the multimodal features include: facial biometric information, human morphological features, including height, head-to-body ratio, and 2D gait temporal features of the human body's external contour; color distribution features, including clothing color histogram features in HSV space, and the feature vector refers to facial biometric information and a structured data set reflecting the macroscopic structure and color distribution of the human body.

[0007] By adopting the above technical solution, multimodal features are extracted and constructed. These multimodal features only circulate within the system. Non-sensitive feature information is extracted from these multimodal features to form descriptive text, which is then used by users to select and identify individuals. This allows the system to avoid sending sensitive privacy data outside the system during the identity recognition process, thus solving the privacy leakage problem of traditional visual monitoring solutions.

[0008] Optionally, the step of non-intrusive identification of the monitored object based on the cluster center and corresponding text description label includes: continuously collecting human feature point information within multiple preset gait cycles; using the DBSCAN clustering algorithm to cluster the feature points, wherein the cluster radius parameter and the minimum number of cluster points parameter are set according to the environmental spatial scale pre-test; extracting only the human morphological features and color distribution features as non-sensitive human feature data from the stable feature clusters formed by clustering, excluding the use of facial biometric information; converting the non-sensitive human feature data into natural language appearance description text and pushing it to the user's mobile terminal for the user to select and confirm the monitored object, thereby completing the non-intrusive identification of the monitored object's identity.

[0009] By adopting the above technical solution, the DBSCAN clustering algorithm is used to cluster feature points within multiple gait cycles. Parameters are set based on environmental spatial scale pre-verification, enabling the system to automatically and accurately divide the feature data of different people into different clusters in a multi-person home environment, thus achieving reliable differentiation of multiple identities. Non-sensitive features are specifically limited to human morphological features and color distribution features, allowing the system to accurately distinguish different individuals based on multiple stable and macroscopic features without relying on sensitive information such as faces, thereby enhancing the robustness of non-sensory labeling.

[0010] Optionally, a target re-identification mechanism is also included: when the monitored object is lost in the field of vision, the multimodal raw data of the candidate target in the current scene is extracted, and the cosine similarity between the target and the center of the feature cluster before the loss is calculated; when the cosine similarity is greater than a preset threshold, the identity of the monitored object is automatically restored.

[0011] By adopting the above technical solution, a target re-identification mechanism is added. After the target is lost, the cosine similarity between the candidate target and the center of the feature cluster before the loss is calculated. This allows the system to automatically restore the identification of the elderly person when they reappear after a short period of time away from the field of vision (e.g., entering the bathroom), thus avoiding monitoring interruption or the need for user reconfirmation.

[0012] Optionally, the optimal observation pose is determined using the following scoring model: ;in, For the total score, The observation effectiveness score is determined by the robot's distance and yaw angle relative to the monitored object; The stealth and undetectable score is determined by the degree of robot exposure within the field of vision of the monitored object; α and β are the corresponding weight coefficients; the method dynamically adapts the weight coefficients according to the monitoring mode: in the private rest mode, the weight of β is increased to achieve covert monitoring; in the emergency response mode, α is set to the maximum weight to achieve close-range real-time observation.

[0013] By adopting the above technical solution, a scoring model with two variable weights, observation effectiveness and stealth and imperceptibility, was constructed for the calculation of optimal observation pose. These weight coefficients were dynamically adapted according to the monitoring mode, which enabled the system to enhance concealment in private rest mode and prioritize close-range real-time observation in emergency response mode. This achieved flexible adaptation to the needs of different scenarios and balanced the daily imperceptible experience with the effectiveness of emergency monitoring.

[0014] Optionally, the distribution logic of the semantic stealth cost layer is as follows: the area within a preset angle range directly in front of the center of the monitored object's line of sight is set as a high-cost area, and the area within a preset angle range outside the high-cost area is set as a second-highest-cost buffer zone; the area behind obstacles in the environmental map is identified, and the space in the blind spot is marked as a gain area according to the position of the monitored object, so as to guide the robot to move along the edge of the shelter.

[0015] By adopting the above technical solution, the distribution logic of the semantic stealth cost layer is limited. The field of vision of the monitored object is divided into high-cost areas and second-highest-cost buffer zones with different cost levels. The blind spots behind obstacles such as furniture are marked as shelter areas with gains. This allows the path planning algorithm to evaluate the environment more precisely and actively guide the robot to choose more concealed routes such as along the edge of the shelter, which greatly improves the stealth of the movement process.

[0016] Optionally, the navigation control steps include active avoidance logic: calculating the collision time between the robot and the monitored object in real time; when the collision time is less than a preset time threshold, the robot performs active avoidance actions such as deceleration or turning, and maintains a distance between itself and the monitored object that is not less than a preset safety threshold.

[0017] By adopting the above technical solution, the navigation control includes active avoidance logic based on collision time, which enables the robot to predict potential collision risks based on relative distance and speed, and to perform active avoidance when the preset time threshold is reached, rather than reacting passively when the distance is too close. This avoids movement conflicts with the elderly in advance and eliminates the fright caused by the robot's sudden approach.

[0018] Optionally, the navigation control steps also include motion smoothing, specifically: using a Sigmoid velocity decay function to smooth the velocity when approaching the target point; and planning the robot's angular velocity and linear velocity using a seven-segment S-shaped acceleration and deceleration curve to eliminate the physical impact during the motion process.

[0019] By adopting the above technical solution, motion smoothing is added to the navigation control. The Sigmoid function is used for smooth speed constraints, and acceleration and deceleration are planned through a seven-segment S-curve. This makes the robot's start-stop and speed changes extremely smooth and natural, eliminating the physical impact and mechanical noise caused by sudden speed changes, and further reducing the auditory and visual disturbance to the elderly.

[0020] Optionally, the navigation control steps adopt an end-to-end path generation scheme based on VLA architecture: extracting environmental visual instructions through the ViT encoder, combining robot state instructions and natural language instructions including field-of-view avoidance instructions; using the Cross-Attention mechanism of the Transformer backbone network to fuse the above-mentioned environmental visual instructions, robot state instructions and natural language instructions, decoding to generate a continuous execution velocity vector, and directly controlling the drive device to perform stealth movement.

[0021] By adopting the above technical solution and using an end-to-end solution based on VLA architecture, the ViT encoder and Transformer network are used to directly process vision, state and natural language commands. This enables the robot to understand abstract commands such as "view avoidance" through autonomous reasoning of the model without the need for manually divided cost maps and fixed rules, and directly generate smooth execution speed, realizing more flexible and intelligent integrated decision-making from perception to stealth movement.

[0022] The second objective of this application is to provide a home-based elderly care monitoring system based on non-intrusive calibration and pose selection, comprising multiple functional modules configured to execute the aforementioned home-based elderly care monitoring method based on non-intrusive calibration and pose selection. Each functional module includes at least: an acquisition module for acquiring multimodal raw data of the monitoring environment and the monitored object; a calibration module for clustering the multimodal raw data using a clustering algorithm and non-intrusively calibrating the monitored object based on the clustering results; a calculation module for calculating the robot's optimal observation pose based on a dual-constraint model of observation effectiveness and stealth non-intrusive behavior, combined with an environmental semantic map, the real-time pose of the monitored object, and its field of vision; and a navigation control module for constructing a multi-layered cost map including a semantic stealth cost layer and executing navigation control to avoid the monitored object's field of vision based on the optimal observation pose. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the home-based elderly care non-intrusive monitoring method based on non-intrusive calibration and posture screening, as described in this application.

[0024] Figure 2 This is a block diagram of the home-based elderly care non-contact monitoring system based on non-contact calibration and posture screening, as described in this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in a comprehensive and detailed manner below. The described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the technical solutions in this application without creative effort are within the scope of protection of this application.

[0026] like Figure 1 As shown in the figure, this application provides a method for non-intrusive monitoring of home-based elderly care based on non-intrusive calibration and pose screening, including the following steps. S01: Obtain multimodal raw data of the monitoring environment and the monitored object.

[0027] S02: Cluster the multimodal raw data using a clustering algorithm, and perform non-intrusive labeling of the monitored objects based on the cluster centers and corresponding text description labels.

[0028] S03: Based on the dual constraint model of observation effectiveness and stealthless movement, combined with the environmental semantic map, the real-time pose of the monitored object and the field of vision, the optimal observation pose of the robot is calculated.

[0029] S04: Construct a multi-layer cost map containing a semantic stealth cost layer, and perform navigation control to avoid the field of vision of the monitored object based on the optimal observation pose; the semantic stealth cost layer sets the field of vision area of ​​the monitored object as a high-cost area and the environmental cover occlusion area as a gain area.

[0030] like Figure 2 As shown, this embodiment of the application implements the above method by building a home-based elderly care non-intrusive monitoring system. Logically, this system includes a mobile robot carrier based on an embedded computing platform. This mobile robot carrier integrates an acquisition module (including multiple environmental perception sensors), a calibration module, a calculation module, and a navigation control module (including a controllable mobile chassis). The acquisition module acquires multimodal raw data of the monitoring environment; the calibration module uses a clustering algorithm to cluster the multimodal raw data and performs non-intrusive calibration of the monitored object based on the clustering results; the calculation module calculates the robot's optimal observation pose based on a dual-constraint model of observation effectiveness and stealth non-intrusiveness, combined with an environmental semantic map, the real-time pose of the monitored object, and its field of vision; the navigation control module constructs a multi-layered cost map including a semantic stealth cost layer and executes navigation control to avoid the monitored object's field of vision based on the optimal observation pose.

[0031] Understandably, this method mainly includes a non-sensory calibration algorithm for monitored objects, an intelligent selection algorithm for optimal observation pose during stealth, and a non-sensory movement control algorithm for stealth navigation. Before the algorithm is started, preliminary basic operations such as environmental geometry mapping and semantic preprocessing need to be completed. In this preliminary basic step, the raw point cloud, images, and inertial data of the home environment can be collected simultaneously by the LiDAR, RGB-D depth camera, and IMU inertial measurement unit carried by the mobile robot. Based on the open-source graph-optimized SLAM algorithm, the multi-source data is processed to construct an indoor grid map with centimeter-level accuracy. The resolution of the grid map is uniformly set to 5cm×5cm. The spatial coordinates, outline dimensions, and occlusion range of static obstacles such as walls, furniture, and appliances are accurately marked in the grid map to form a basic geometric map. Subsequently, a lightweight vision-language model is introduced. The geometric map undergoes semantic enhancement to automatically identify the specific locations of furniture such as beds, sofas, dining tables, and cabinets, and accordingly divides functional areas into living rooms, dining rooms, bedrooms, and bathrooms. Bedrooms and bathrooms are clearly marked as privacy-sensitive areas, while living rooms and dining rooms are marked as priority observation areas. At the same time, the system identifies and marks the elderly's regular activity routes and safe stealth passages such as those behind furniture. Compliant alternative observation points are preset around the areas where the elderly frequently move, and invalid points in obstructed areas and prohibited areas are eliminated, forming a comprehensive alternative observation point library that is stored locally in the system. The standardized indoor geometric map, semantic region annotation information, static obstacle parameters, data on the elderly's regular activity routes, coordinates of safe stealth passages, and the comprehensive alternative observation point library output by this preliminary step all serve as fixed basic parameters for the operation of all subsequent algorithms.

[0032] Based on the aforementioned preliminary preparations, the real-time perception data acquisition phase proceeds. This phase can be executed at a default adjustable fixed acquisition frequency of 10Hz, serving as the sole standardized data input source for the aforementioned object-monitoring calibration algorithm, stealth optimal observation pose intelligent selection algorithm, and stealth navigation stealth motion control algorithm. The acquired multimodal raw data is divided into two main categories: object-monitoring parameters and environmental dynamic parameters. Regarding object-monitoring parameters, multimodal features are collected and extracted from the multimodal raw data, and feature vectors are constructed. Multimodal features include: facial biometric information, human morphological features, and color distribution features. Human morphological features include height, head-to-body ratio, and human external contour. D. Gait temporal characteristics; color distribution characteristics include clothing color histogram features in HSV space; feature vectors refer to facial biometric information, a structured data set reflecting the macroscopic structure and color distribution of the human body, real-time spatial coordinates, motion velocity vectors, angular velocity, and motion trajectory prediction data of the elderly (a specific embodiment of the monitored object), the elderly's body orientation angle, visual field range defined based on 120° horizontally, and current activity state, which can be subdivided into lying, sitting, walking, and turning, as well as the real-time distance and relative azimuth angle between the elderly and the candidate observation points. In terms of environmental dynamic parameters, the collected content includes the robot's own real-time positioning coordinates and orientation angle. The system collects the real-time location and range of dynamic obstacles in the environment, the semantic attributes of the current scene (identified as a privacy zone or observation zone), and the real-time occlusion status of candidate points. After collection, all parameters are integrated into standardized data frames. Simultaneously, based on the elderly person's status and environmental semantic parameters, the system automatically determines the scene mode, classifying the current scene into three categories: regular monitoring mode, private rest mode, and emergency / abnormal mode. The determination rule is that the system defaults to regular monitoring mode when the following two special mode conditions are not met: the elderly person is in a non-privacy observation priority area such as the living room or dining room, their activity is normal walking, sitting, or leisure activities, and there are no abnormal alarm signals. The private rest mode is triggered when the elderly person is in a privacy-sensitive area such as a bedroom or bathroom and remains lying or sitting still without making large limb movements for more than 30 seconds, without any abnormal alarm signals such as falling or prolonged stillness. The emergency abnormal mode is immediately and forcibly triggered when the system detects monitoring alarm signals such as the elderly person falling, suddenly remaining still for a long time without movement, abnormal posture, or calling for help. The mode determination result is pushed synchronously with the standardized data frame to the aforementioned monitoring object non-sensory calibration algorithm, stealth optimal observation posture intelligent selection algorithm, and stealth navigation non-sensory movement control algorithm as the basis for subsequent weight adaptive adjustment, ensuring that the parameters obtained by each algorithm are completely consistent, without delay, and without omission.

[0033] After acquiring real-time perception data across the entire domain, the first step is to execute a non-intrusive identification algorithm for the monitored object. The non-intrusive identification steps for the monitored object based on the cluster center and corresponding text description labels include: continuously collecting human feature point information within multiple preset gait cycles; using the DBSCAN clustering algorithm to cluster the feature points, wherein the cluster radius parameter and the minimum number of cluster points parameter are pre-set according to the environmental spatial scale; extracting only the human morphological features and color distribution features from the stable feature clusters formed by clustering as non-sensitive human feature data, excluding the use of facial biometric information; converting the non-sensitive human feature data into natural language appearance description text and pushing it to the user's mobile terminal for the user to select and confirm the monitored object, thus completing the non-intrusive identification of the monitored object's identity.

[0034] Understandably, the following logic can be adopted during the execution process: First, feature extraction is performed. From the perceived data, features including facial biometric information, height range, head-to-body ratio, and 2D human body contour gait temporal features of 3 to 5 consecutive cycles are extracted, taking the elderly person's single leg from leaving the ground, taking a step, and landing back down as one complete walking cycle. Clothing color features are also extracted based on the HSV color space to generate a fixed-dimensional feature vector. The extraction rule for HSV clothing color features is to extract only the hue mean of the non-facial area of ​​the elderly person's torso to automatically filter out interference such as clothing accessories and lace, lock the basic color tone of the main clothing, extract the saturation mean to eliminate color deviations caused by ambient light and stains, thereby improving feature stability, and extract the brightness mean to adapt to the lighting differences in different areas such as indoor living rooms and bedrooms, avoiding color recognition failure caused by light intensity. This set of HSV features only extracts the quantized mean values ​​of the three major channels of hue, saturation, and brightness, and combines them into a three-dimensional color feature. The original image of the clothing is neither stored nor transmitted back. Subsequently, the four features—height range, head-to-body ratio, gait timing, and HSV clothing color—were standardized and quantized to fix the corresponding dimensions of each feature and eliminate redundant and invalid parameters, ensuring a constant and uniform number of parameters for each feature. Then, all quantized feature values ​​were uniformly mapped to the [0,1] interval for dimensional normalization, eliminating differences in numerical magnitude and dimensions between different features to avoid excessively high weights of a single feature interfering with the accuracy of subsequent clustering algorithms. Finally, the normalized feature parameters were concatenated in a fixed order—height range, head-to-body ratio, gait timing, and HSV clothing color—to generate a multidimensional structured feature vector with constant dimensions and a regular format. This feature vector is characterized by a constant total dimension and a uniform data format, remaining unchanged regardless of the data collection scenario or individual differences.

[0035] After extracting the feature vectors, the multi-target density clustering differentiation stage begins. The open-source DBSCAN density clustering algorithm can be used to cluster all feature vectors collected within 24 hours. Based on the actual spatial scale of the home environment and the prior knowledge of the presence of a small number of family members, the minimum number of samples per cluster can be set to 5, and the maximum neighborhood distance can be set to 0.3. This parameter setting automatically distinguishes different family members, ensuring that each cluster corresponds to a single family member, thus avoiding multi-target confusion that could interfere with subsequent monitoring procedures. As the clustering process completes, only the human data is extracted from the stable feature clusters formed. Body shape features and color distribution features are treated as non-sensitive human feature data, excluding the use of facial biometric information. The non-sensitive human feature data is used to generate natural language appearance description text through semantic matching. For example, a natural language appearance description such as "an elderly person who is about 1.6 meters tall, walks slowly, and usually wears dark clothes" is generated and pushed to the user's mobile app or web interface. The user can confirm the person to be monitored with a one-click checkbox, thereby completing the non-intrusive identification of the person to be monitored. After the identification is completed, the system only retains the feature vector library locally and immediately destroys the original image and feature collection data, further strengthening the protection of privacy.

[0036] It is understood that the method in this application embodiment also includes a target re-identification mechanism: when the monitored object is lost in the field of vision, the feature vector of the candidate target in the current scene is extracted, and the cosine similarity between it and the center of the feature cluster before the loss is calculated; when the cosine similarity is greater than a preset threshold, the identity of the monitored object is automatically restored. The specific execution process of this target re-identification mechanism is as follows: when the elderly person leaves the sensor's perception range, such as entering a private area like a bathroom, the system automatically records the spatial coordinates of the elderly person's disappearance and the motion vector before disappearance. After the elderly person re-enters the perception range, the open-source cosine similarity algorithm is used to compare the feature vector extracted in real time with the vector of the labeled feature library. The similarity threshold can be set to 0.85. It is understood that this value is the optimal threshold selected by taking into account the recognition accuracy and anti-interference of gait and clothing features. Once the similarity is greater than or equal to 0.85, it is determined to be the same monitored target, and the system automatically restores its identity labeling state and restarts the monitoring process. No secondary operation is required from the user throughout the process, effectively ensuring the continuity and integrity of the monitoring.

[0037] After target locking is completed, the system performs a selection calculation for the optimal observation pose based on a dual constraint model of observation effectiveness and stealthless operation. The optimal observation pose is determined through the following scoring model: ;in, For the total score, The observation effectiveness score is determined by the robot's distance and yaw angle relative to the monitored object; The stealth-free score is determined by the degree of robot exposure within the monitored subject's field of vision; α and β are the corresponding weighting coefficients. During the calculation, a preliminary screening of candidate points is first performed. Centered on the real-time spatial coordinates of the target elderly person calibrated by the monitored subject's stealth-free calibration algorithm, polar coordinate sampling is performed within the effective observation distance range [1.5m, 5m]. Combined with the environmental parameters obtained from the previous steps, points that are obscured, points in privacy areas, and points located on the elderly person's regular activity lines are eliminated, leaving only the set of compliant candidate points. Subsequently, the comprehensive score of each point in the candidate set is calculated to achieve an equal balance between observation effectiveness and stealth-free operation.

[0038] When constructing the observation effectiveness score, it is composed of two weighted components: visibility score and distance adaptation score. The visibility score has a weight of 0.6, and a perfect score is awarded when the line of sight is completely unobstructed; if there is obstruction, points are deducted accordingly based on the obstruction ratio. The distance adaptation score has a weight of 0.4, and within the effective observation distance range of 1.5 meters to 5 meters, 3 meters is set as the optimal observation distance and awarded a perfect score for this item. When constructing the stealth-undetectable score, it is composed of two weighted components: field of view deviation score and cover obstruction score. The score for visual field deviation accounts for 0.5 of the total score. The calculation logic is that the further the robot deviates from the elderly person's 120° main visual field, the higher the score. The score for occlusion also accounts for 0.5 of the total score. The calculation logic is that the more obstructions present near the robot, the higher the score. Here, occlusion does not refer to simple visual obstruction, but rather to the degree to which physical obstructions such as furniture and cabinets in the environment cover the robot. Then, based on the scenario-based model and combined with the collected real-time state parameters of the elderly person, the weight coefficients for observation effectiveness and stealthless operation are adaptively adjusted. In the protection mode, the system maintains an equal-drive approach, keeping the weight values ​​at α=0.5 and β=0.5 to balance observation effectiveness and a seamless monitoring experience. In the private rest mode, priority is given to ensuring seamless concealment and privacy protection; for example, the weight values ​​are adjusted to β=0.7 and α=0.3 to weaken the observation weight and strengthen the stealthy and unobtrusive weight. In the emergency anomaly mode, the system switches to an observation-priority mechanism to fully ensure the monitoring effectiveness of abnormal states; for example, the weight values ​​are forcibly adjusted to α=1.0 and β=0. Once the abnormal signal is resolved and the elderly person resumes normal activities, the system automatically switches back to the regular monitoring mode. The weight configuration is such that the weight coefficients are dynamically adapted according to the monitoring mode: in the private rest mode, the weight of β is increased to achieve covert monitoring; in the emergency response mode, α is set to the maximum weight to achieve close-range real-time observation. After the calculation is completed, the score of the selected point is evaluated in real time. If the total score is lower than 0.6, the re-screening logic is triggered. An exploration-utilization mechanism is introduced for high-frequency dwell points. In 80% of the cases, the historical best point is used and in 20% of the cases, the calculation is performed to prevent getting stuck in a local optimum. The point library is continuously optimized through long-term operation and iteration.

[0039] After obtaining the optimal observation pose point array and its included coordinates and orientation angle information output by the stealth optimal observation pose intelligent selection algorithm, the stealth navigation seamless movement control algorithm enters the execution stage. This stage constructs a multi-layered cost map containing a semantic stealth cost layer. Based on the optimal observation pose, navigation control is executed to avoid the monitored object's field of vision. The semantic stealth cost layer sets the monitored object's field of vision area as a high-cost zone and the environmental cover occlusion area as a gain zone. The construction of the multi-layered cost map is based on real-time environmental parameters and semantic annotation information collected by the global real-time perception module. An additional semantic stealth cost layer is superimposed on the basic obstacle cost layer. The robot's path generation is guided by quantifying cost values ​​to achieve autonomous path avoidance of highly disturbing areas. The cost value here is a quantified value specifically for robot path planning; a higher value indicates a greater risk of disturbing or interfering with the elderly person when traversing that area. The algorithm automatically avoids high-cost areas and prioritizes low-cost areas. The specific cost partitioning and assignment logic is as follows: The area within a preset angle directly in front of the monitored person's line of sight is designated as a high-cost zone, and the area within a preset angle outside the high-cost zone is designated as a second-highest-cost buffer zone. The area behind obstacles in the environmental map is identified, and spaces in blind spots are marked as gain zones based on the monitored person's position to guide the robot to move along the edge of the cover. In the specific quantification and assignment operation, the 30° field of vision area directly in front of the elderly person and the 120° horizontal main line of sight sector are designated as high-cost zones and assigned a cost value of 9. The cost value of this area increases inversely with the square of the distance from the elderly person. The algorithm directly avoids and prohibits the robot from entering this area. The area of ​​the elderly person's daily activity path is designated as a medium-cost zone and assigned a cost value of 4. The algorithm tries to detour around this area to reduce obstruction to the elderly person's normal activities. Concealed cover areas such as behind furniture, cabinets, and sofas are designated as low-cost or gain zones. The weight of this area can be set to a negative value, assigned a cost value of 1 or lower, to guide the algorithm to prioritize the planning path in this area, thereby maximizing the reduction of visual disturbance.

[0040] The navigation control steps based on multi-layer cost maps include active avoidance logic, specifically calculating the time-to-collision (TTC) between the robot and the monitored object in real time. When the TTC is less than a preset time threshold, the robot performs active avoidance actions such as deceleration or turning, maintaining a distance from the monitored object that is not less than a preset safety threshold. This active predictive avoidance logic completely abandons the traditional robot's reactive operation mode of passively starting, stopping, or detouring only after detecting an obstacle or a person approaching. Instead, it adopts a motion vector-based predictive approach to proactively avoid potential disturbance risks. The execution logic is as follows: The robot collects the elderly person's real-time motion vector, which specifically includes linear velocity and angular velocity, the robot's own real-time motion parameters, and the relative spatial position parameters between the two. First, the TTC is calculated. The specific formula for calculating TTC is... ,in The variable represents the real-time relative straight-line distance between the robot and the elderly person, in meters, and is collected in real time by the global perception module. The variable represents the real-time relative approach speed between the robot and the elderly person, measured in meters per second. This relative approach speed is calculated by the vector difference between the elderly person's real-time linear velocity and the robot's real-time linear velocity. During the calculation, only the effective velocity components of the two approaching each other are calculated, and invalid components such as parallel and opposite movements are eliminated to avoid the risk of misjudgment. Based on the above calculation results, the robot accurately predicts the elderly person's movement trajectory for the next 2 seconds. When the calculated TTC is less than or equal to a preset time threshold of 1.5 seconds, it is immediately determined that there is a potential risk of disturbance. Then, without waiting for the elderly person to enter the avoidance threshold, the robot initiates an avoidance action in advance. Throughout the avoidance process, the robot maintains a minimum safe distance of 0.8 meters from the elderly person. The avoidance direction is preferentially selected to be an area outside the predicted movement route of the elderly person, and the avoidance destination is preferentially locked to the low-cost safe sneak area marked in the multi-layer cost map, thereby ensuring that the avoidance action itself will not reverse into the elderly person's activity range.

[0041] The navigation control process also includes motion smoothing, specifically: using a Sigmoid velocity decay function to smooth the velocity when approaching the target point; and planning the robot's angular and linear velocities using a seven-segment S-shaped acceleration and deceleration curve to eliminate physical impact during movement. This refined, seamless motion control strategy addresses the physiological sensitivity of the elderly or their susceptibility to sudden movements, dynamically regulating the entire robot movement process to completely eliminate abruptness and mechanical vibration caused by sudden starts, stops, sharp turns, or speed changes. Regarding dynamic speed adaptation, the movement speed is dynamically adjusted based on the relative distance between the robot and the elderly person, collected by the sensing module, and the direction of the elderly person's attention. When the distance between the robot and the elderly person is greater than 3 meters, a maximum speed of 0.5 m / s is allowed. When the distance is between 1.5 meters and 3 meters, the velocity v can gradually decay according to a Sigmoid smoothing nonlinear function. In this scenario, the Sigmoid function is specifically expressed as follows: Where d represents the real-time distance between the robot and the elderly person, this function belongs to the general smooth transition function in the field of robot motion control. Its role in this application is to realize the nonlinear and gradual transition of the robot from high speed at a long distance to low speed at a short distance, replacing the linear speed decay, thereby completely eliminating the mechanical vibration and visual impact caused by sudden speed changes. When the global perception module detects that the elderly person's gaze is focused on the robot, that is, when the attention turns to the device, the system immediately triggers an instantaneous stop command to reduce the speed to zero. After the elderly person's attention shifts, the movement is resumed smoothly from zero speed. In terms of smooth motion acceleration and deceleration control, an open-source seven-segment S-shaped acceleration and deceleration curve can be used to limit the acceleration. For example, the maximum acceleration can be limited to 0.2 m / s². 2 Within this range, the jerk will be limited to 0.1 m / s². 3 Within this range, the start-stop and acceleration / deceleration processes are smooth and seamless. All turning actions utilize large-radius Bézier curves for path planning to avoid sharp turns with small radii, while simultaneously reducing mechanical friction noise to minimize auditory interference for the elderly. Furthermore, an intelligent timing selection strategy for movement can be introduced. The system prioritizes moving actions during periods when the elderly are focused, such as while watching TV, reading, or sitting quietly and their gaze is not focused on the robot. During movement, the system monitors changes in the elderly's gaze and attention in real time. If the system detects that the elderly are paying attention to a device, it immediately pauses movement and remains stationary until the elderly resume normal activity and their attention shifts, thus minimizing the probability of being detected.

[0042] Regarding path search algorithm adaptation, it supports the open-source A* algorithm and the classic Dijkstra path search algorithm for optimal path search on multi-layered cost maps that have already been superimposed. An optimization objective function is set, aiming to minimize the total cost while considering path length, stealth and concealment, and operational smoothness. The specific formula is as follows: ,in, This represents the total stealth cost of the path and is a metric used by the algorithm to select paths. The smaller the value, the lower the overall interference of the path and the better the stealth performance. The algorithm ultimately selects the path with the minimum total cost as the execution route. This is the path length weight coefficient, with a value of 0.2. It is a manually preset adjustable hyperparameter used to balance the impact of path length on the total cost. Its value follows the principle of "stealth priority is higher than path length," ensuring that the weight value is lower than the stealth cost weight to prevent the algorithm from entering high-disturbance regions in order to shorten the path. Represents the actual physical length of the planned path, in meters. The stealth and concealment weight coefficient is set to 0.6. As a weighting coefficient, its value is much higher than other weights, thus highlighting the algorithm's principle of "unobtrusive avoidance and low disturbance." This directly corresponds to the assigned values ​​in the multi-layered cost map based on areas such as the field of view, movement paths, and concealed cover. Higher values ​​indicate a greater risk of disturbing the elderly along that path. This is the motion smoothness weighting coefficient, with a value of 0.2, used to adjust the impact of robot motion smoothness on the total cost. The total deviation of the path turning angle represents the cumulative change in the angle of all turning actions during the robot's movement along the planned path. The smaller the value, the smoother the path and the fewer the turns, which can effectively avoid mechanical noise and visual abruptness caused by sharp turns. Through the dynamic weighted calculation of this objective function, each weight can be finely adjusted within ±0.1 according to the scenario to achieve the optimal balance between path length, stealth and concealment, and smoothness of movement, thus implementing the principle that "disturbance avoidance has a higher priority than operational stability, and operational stability has a higher priority than path length".

[0043] The navigation control steps can also employ an end-to-end path generation scheme based on a VLA architecture: Environmental visual instructions are extracted through a ViT encoder, combined with robot state instructions and natural language instructions including field-of-view avoidance commands; the Cross-Attention mechanism of the Transformer backbone network is used to fuse these environmental visual instructions, robot state instructions, and natural language instructions, decoding to generate continuous execution velocity vectors, directly controlling the drive device to perform stealth movement. This advanced extended route uses a vision-language-action multimodal large model to achieve end-to-end path generation and motion control, eliminating the need for manual cost assignment and constraint logic decomposition. It achieves more flexible scene adaptation entirely through autonomous model reasoning. When executing this architecture, it can be implemented using an industry-mature... Using the open-source VLA vision-language-action multimodal pre-trained model as the initialization base, this framework reuses its mature Transformer backbone, ViT visual encoder, and cross-modal attention infrastructure. Based on this general architecture, it is specifically fine-tuned for the home-based elderly care seamless stealth scenario. The entire VLA multimodal end-to-end inference framework includes a perception encoding module, a semantic inference module, and an action decoder. The perception encoding module, as the front-end perception of the multimodal large model, follows the data preprocessing and feature fusion principles of multimodal large models. It is responsible for standardizing multi-source heterogeneous perception data, unifying modal representation, and deep fusion. It extracts environmental RGB-D visual features through the ViT visual encoder to generate visual instructions. ViT stands for Vision. The Transformer is a dedicated visual feature extraction backbone in the VLA multimodal large model. It relies on a self-attention mechanism to capture global visual information rather than extracting local features. Specifically designed for home scenarios, it processes RGB-D depth visual data to simultaneously consider environmental color and texture information as well as spatial depth and distance information. This allows for accurate identification of semantic targets such as elderly contours, furniture positions, spatial boundaries, and visual field regions. It transforms unstructured visual data into structured features to meet the standardized input requirements of the multimodal large model for visual modalities. During the generation of visual instructions, the ViT encoder segments the complete RGB-D environmental image into fixed-size visual image blocks. Each image block undergoes linear mapping and positional encoding. The data is then transformed into independent visual commands, essentially decomposing unstructured environmental visual information into standardized discrete feature units recognizable by the multimodal model. Each visual command carries the visual features and spatial location information of the corresponding local area. All commands are combined to form a complete global environmental visual feature sequence, facilitating cross-modal global correlation reasoning in the subsequent Transformer module. This avoids missing key visual information such as the elderly person's activity area and high-cost disturbance areas. Simultaneously, the real-time position, posture, and motion vectors of the elderly person, as well as the robot's own localization and posture information, are encoded to generate state commands. The generation of state commands follows the principle of unified representation of heterogeneous data in multimodal large models to solve the problem of inconsistent dimensions and formats of data from different modalities, which prevent direct fusion.Its data sources include the elderly person's real-time 3D coordinates, body postures (sitting, lying, standing, walking), movement speed and direction vectors, as well as dynamic data such as the robot's own positioning coordinates, chassis posture, and movement speed. First, various heterogeneous state data undergo normalization preprocessing to unify data units and numerical ranges, eliminating numerical differences between different dimensions. Then, a dedicated fully connected layer maps the normalized state data into feature vectors—state commands—that are completely consistent with the visual command dimension. Finally, the visual commands and state commands are concatenated and fused along the feature dimension, achieving unified encoding of multi-source data from the visual environment, human body state, and the robot's own state, forming a complete multimodal fusion feature sequence. This lays the foundation for subsequent cross-modal alignment with language constraints, enhancing semantics. The reasoning module enables autonomous understanding of stealth logic without conscious awareness. It introduces a cross-attention layer that precisely aligns natural language instructions, including explicit visual avoidance and low-interference stealth constraints, such as "Do not enter the high-cost visual area directly in front of the elderly," "Prioritize moving to concealed areas behind furniture," "Maintain a minimum safe distance of 0.8 meters from the elderly," and "No sudden turns or abrupt starts / stops during movement," with visual instructions. The cross-attention layer, a feature of the Transformer architecture, differs from the single-sequence feature association of the self-attention mechanism. It enables dynamic association and weight matching between two different feature sequences. In this module, the cross-attention layer is responsible for associating the linguistic constraint feature sequence with the visual instruction sequence through calculation. The attention weighting between linguistic and visual features automatically identifies regions in the environmental visual features corresponding to linguistic constraints, accurately matching, for example, "elderly visual field" with corresponding visual instructions, and "concealed lurking area" with corresponding visual instructions. This achieves dynamic binding between linguistic rules and visual scenes. The precise alignment logic relies on a cross-attention layer to achieve bidirectional alignment. On one hand, the linguistic feature sequence transformed from linguistic constraint prompts is located to the corresponding key region features in the visual instructions through attention weight allocation, marking high-disturbance areas, safe lurking areas, and elderly activity areas. On the other hand, the environmental features of the visual instructions are fed back to the linguistic features, allowing the model to adjust the constraint enforcement strength according to the real-time environment. This enables the model to autonomously internalize unconscious lurking rules without requiring manual setting of hierarchical costs. To enable flexible constraint reasoning in complex and ever-changing home scenarios, the action decoder, as the model output terminal, is responsible for converting the high-level semantic features after fusion reasoning into low-level motion commands that the robot chassis can directly execute. This is achieved through a multilayer perceptron, which outputs continuous and smooth motion control vectors, namely linear velocity and angular velocity. The multilayer perceptron is a classic feedforward artificial neural network, composed of multiple layers of fully connected neurons, including an input layer, multiple hidden layers, and an output layer. It possesses powerful nonlinear feature mapping and fitting capabilities. Its input is the fused feature vector output by the semantic reasoning module. After nonlinear transformation and feature extraction through multiple hidden layers, two continuous values ​​are finally generated at the output layer, corresponding to the linear velocity and angular velocity of the robot chassis, respectively. Furthermore, this design is specifically tailored to the requirement of seamless stealth operation.By adding a smoothing constraint to the multilayer perceptron, the output control vector becomes a continuously and gradually changing value to prevent abrupt changes. This ensures smooth acceleration, deceleration, and turning movements of the robot and directly generates executable chassis control commands, thus eliminating the post-processing step required in traditional algorithms.

[0044] The analysis phase training process of this VLA multimodal end-to-end inference framework can be divided into two stages. The first stage is basic pre-training. After initializing the weights of the open-source VLA multimodal pre-trained model, lightweight basic transfer training is completed on a general indoor navigation public dataset. This enables the model to learn basic navigation capabilities such as indoor mapping, dynamic obstacle avoidance, autonomous localization, and chassis closed-loop control, thereby solidifying the underlying motion stability and general environment adaptability. This avoids the need for training a large model base from scratch, thus balancing R&D efficiency and computing costs. The second stage is stealth-specific fine-tuning. Using a home-based elderly care scenario dataset with refined disturbance scoring labeled using the method in the embodiments of this application, combined with supervised fine-tuning and reinforcement learning for joint optimization, the reward function value is set to be strictly negatively correlated with the probability of the elderly being noticed and the degree of disturbance. This guides the model to autonomously learn low-disturbance scenarios suitable for home-based elderly care. The low-disturbance stealth path planning and smooth motion control strategy fine-tunes the labeled dimensions of the dataset, which include three categories of labels: disturbance level in the home scene, visual avoidance area, and stealth path compliance. The input-output mapping relationship is as follows: the input end integrates the current frame environmental image, semantic map, real-time status and intention of the elderly, and dual constraints of observation effectiveness and stealth non-perceptible operation obtained by the global perception module; the output end generates the optimal linear velocity and angular velocity motion vector for the next 0.5 seconds, realizing the integrated output of path planning and motion control. Although this technical approach abandons the manual rule-based cost assignment, it still needs to follow the constraints of observation effectiveness and stealth non-perceptible operation. The path and motion commands generated by the model inference must not only ensure that the optimal observation point is reached for effective monitoring, but also meet the requirements of visual avoidance, no disturbance, and low interference throughout the process, ensuring that the criteria of the two technical approaches are consistent.

[0045] This application embodiment can simultaneously support both rule-based stealth control and VLA multimodal large model end-to-end stealth. This dual technical approach can be flexibly switched according to hardware computing power and scenario requirements. For embedded platform systems with computing power less than or equal to 8 TOPS, the first approach, i.e., deterministic rule-based stealth, is automatically locked to achieve efficient and stable operation. For high-performance platforms with computing power greater than 8 TOPS, the second approach, i.e., VLA multimodal large model stealth, is supported to achieve more intelligent and flexible end-to-end control. The mode switching can be manually locked or automatically identified by the system. The two modes cannot run in parallel after switching. However, regardless of which technical approach is adopted, the principles of observation effectiveness constraint and stealth non-sensory constraint must be followed. The paths and motion commands generated by both approaches must simultaneously meet the rigid requirements of reaching the optimal observation point and avoiding obstacles throughout the entire field of view to ensure that the final effect of non-sensory stealth remains consistent.

[0046] The final output of the stealth navigation and seamless movement control algorithm is a smooth and continuous chassis motion control command, driving the robot's mobile platform to move along the planned path and accurately reach the optimal observation point output by the stealth optimal observation pose intelligent selection algorithm. Upon arrival, the system immediately switches to a silent fixed-point observation mode. The robot remains stationary throughout, without generating mechanical vibrations, flashing lights, or audible prompts—signals that might attract the elderly's attention. Simultaneously, the full-domain real-time perception module is activated to selectively collect essential monitoring data such as the elderly person's body posture, real-time location, activity range, and movement trends. This ensures continuous, seamless observation. At the same time, the system initiates a fully automatic dynamic closed-loop iterative update mechanism to adapt to the elderly person's autonomous activities and the dynamic changes in the indoor environment. Once preset observation failure or optimization conditions are triggered, specifically including changes in the elderly person's position... If the movement exceeds the preset optimal observation effect threshold, causing the original position to lose its best observation angle; if indoor furniture is moved or clutter is piled up, obstructing the observation line of sight and preventing effective monitoring; if the elderly person's gaze is continuously focused on the robot, thus violating the non-intrusive stealth constraint and requiring a change of concealed observation position; or if the elderly person's posture or activity state suddenly changes, requiring adjustment of the position to ensure comprehensive monitoring, the system automatically interrupts the current fixed-point observation task and keeps the robot in a silent standby state. Then, it immediately backtracks to the optimal observation position, the calculation module recalculates the optimal observation position adapted to the current scene, and after the calculation is completed, the stealth navigation algorithm generates a new non-intrusive stealth path. The robot then performs low-disturbance stealth again and restarts fixed-point observation after reaching the new position. This forms a seamless and fully automatic closed-loop process. The entire closed-loop operation can adapt to dynamic scene changes without human intervention.

[0047] During the closed-loop operation of the entire process, the privacy protection mechanism bound to the algorithm is forcibly activated. This privacy protection mechanism is deeply linked with the algorithms of all-domain perception, stealth navigation, and fixed-point observation, and cannot be skipped or turned off individually. It builds a strong privacy defense line from data collection, processing to storage. Specifically, it is reflected in two dimensions: localized privacy processing mechanism and multi-person scene adaptive privacy protection mechanism. At the localized processing mechanism level, all monitoring-related data collected by the perception unit is analyzed, calculated, and judged in real time on the device's local main control processing unit. It is not transmitted to the outside world, not uploaded to cloud storage, and does not disclose the original collected data. It only outputs the elderly's activity status, abnormal posture, and position deviation. These non-privacy analysis results and monitoring warning signals aim to eliminate the risk of privacy leaks from the data processing perspective. At the level of adaptive privacy protection mechanism in multi-person scenarios, when the all-domain perception unit detects irrelevant personnel other than the target elderly person being monitored within the observation range, the system will instantly link up through algorithms and immediately suspend all monitoring data collection and observation analysis actions. The robot will remain silent and still, without collecting, processing, or storing any scene-related data. After the irrelevant personnel have completely left the effective observation range, the system will automatically resume normal fixed-point observation tasks. The entire switching process is seamless and without any additional actions, thereby avoiding the risk of privacy leaks in multi-person home scenarios while taking into account the practicality and compliance of monitoring.

[0048] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0049] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0050] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0051] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0052] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0053] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for non-intrusive monitoring of home-based elderly care based on non-intrusive calibration and pose screening, characterized in that, include: Acquire multimodal raw data of the monitoring environment and the monitored individuals; Clustering algorithms are used to cluster the multimodal raw data, and the monitored objects are non-intrusively labeled according to the cluster centers and corresponding text description labels. Based on the dual constraint model of observation effectiveness and stealthless movement, combined with the environmental semantic map, the real-time pose of the monitored object and the field of vision, the optimal observation pose of the robot is calculated. A multi-layer cost map containing a semantic stealth cost layer is constructed. Based on the optimal observation pose, navigation control is executed to avoid the field of vision of the monitored object. The semantic stealth cost layer sets the field of vision area of ​​the monitored object as a high-cost area and the environmental cover occlusion area as a gain area.

2. The home-based elderly care non-contact monitoring method based on non-contact calibration and pose screening according to claim 1, characterized in that, Also includes: Extract multimodal features from the original multimodal data and construct feature vectors; The multimodal features include: facial biometric information, human morphological features, including height, head-to-body ratio, and 2D gait temporal features of the human body outline; Color distribution features include clothing color histogram features in HSV space. The feature vector refers to facial biometric information and a structured data set that reflects the macroscopic structure and color distribution of the human body.

3. The home-based elderly care non-contact monitoring method based on non-contact calibration and pose screening according to claim 2, characterized in that, The step of non-intrusive identification of the monitored object based on the cluster center and the corresponding text description label includes: The system continuously collects human feature point information within multiple preset gait cycles. The DBSCAN clustering algorithm is used to cluster the feature points. The clustering radius parameter and the minimum number of cluster points parameter are set according to the environmental spatial scale prediction. From the stable feature clusters formed by clustering, only the human morphological features and color distribution features are extracted as non-sensitive human feature data, excluding the use of facial biometric information; the non-sensitive human feature data is converted into natural language appearance description text and pushed to the user's mobile terminal so that the user can select and confirm the subject to be monitored, thus completing the seamless identification of the subject's identity.

4. The home-based elderly care non-contact monitoring method based on non-contact calibration and pose screening according to claim 1, characterized in that, It also includes a target re-identification mechanism: When the monitored object is lost from the field of vision, the multimodal raw data of the candidate target in the current scene is extracted, and its cosine similarity with the center of the feature cluster before the loss is calculated. When the cosine similarity is greater than a preset threshold, the identity of the monitored object is automatically restored.

5. The home-based elderly care non-contact monitoring method based on non-contact calibration and pose screening according to claim 1, characterized in that, The optimal observation pose is determined using the following scoring model: ;in, For the total score, The observation effectiveness score is determined by the robot's distance and yaw angle relative to the monitored object; The stealth and undetectable score is determined by the degree of robot exposure within the field of vision of the monitored object; α and β are the corresponding weight coefficients; the method dynamically adapts the weight coefficients according to the monitoring mode: in the private rest mode, the weight of β is increased to achieve covert monitoring; in the emergency response mode, α is set to the maximum weight to achieve close-range real-time observation.

6. The home-based elderly care non-contact monitoring method based on non-contact calibration and pose screening according to claim 1, characterized in that, The distribution logic of the semantic stealth cost layer is as follows: The area within a preset angle directly in front of the center of the monitored object's line of sight is designated as the high-cost zone, and the area within a preset angle outside the high-cost zone is designated as the second-highest-cost buffer zone. Identify the area behind obstacles in the environmental map, and mark the space in the blind spot as a gain zone according to the position of the monitored object, so as to guide the robot to move along the edge of the shelter.

7. The home-based elderly care non-contact monitoring method based on non-contact calibration and pose screening according to claim 1, characterized in that, The navigation control steps include active obstacle avoidance logic: Real-time calculation of the collision time between the robot and the monitored object; When the collision time is less than a preset time threshold, the robot performs active avoidance actions such as deceleration or turning, and maintains a distance from the monitored object that is not less than a preset safety threshold.

8. The home-based elderly care non-contact monitoring method based on non-contact calibration and pose screening according to claim 1, characterized in that, The navigation control steps also include motion smoothing processing, specifically: The sigmoid velocity decay function is used to smoothly constrain the velocity when approaching the target point; The robot's angular velocity and linear velocity are planned using a seven-segment S-shaped acceleration and deceleration curve to eliminate the physical impact during motion.

9. The home-based elderly care non-contact monitoring method based on non-contact calibration and pose screening according to claim 1, characterized in that, The navigation control steps employ an end-to-end path generation scheme based on a VLA architecture: The ViT encoder extracts environmental visual instructions, which are then combined with robot status instructions and natural language instructions that include field-of-view avoidance instructions. By utilizing the Cross-Attention mechanism of the Transformer backbone network to fuse the aforementioned environmental visual instructions, robot state instructions, and natural language instructions, a continuous execution velocity vector is generated through decoding, which directly controls the drive device to perform stealth movement.

10. A home-based elderly care monitoring system based on non-contact calibration and pose screening, characterized in that, The system includes multiple functional modules configured to perform the home-based elderly care non-contact monitoring method based on non-contact calibration and pose screening as described in any one of claims 1 to 9, wherein the functional modules include at least: The acquisition module is used to acquire multimodal raw data of the monitoring environment and the monitored object; The calibration module is used to cluster the multimodal raw data using a clustering algorithm and to perform non-sensory calibration of the monitored object based on the clustering results. The calculation module is used to calculate the robot's optimal observation pose based on the dual constraint model of observation effectiveness and stealthless movement, combined with the environmental semantic map, the real-time pose of the monitored object, and the field of vision. The navigation control module is used to construct a multi-layer cost map containing a semantic stealth cost layer, and to perform navigation control to avoid the field of vision of the monitored object based on the optimal observation pose.

Images