Self-walking path recognition and planning method and system for lower limb rehabilitation robot

Abstract

This application provides a method and system for autonomous walking path recognition and planning for a lower limb rehabilitation robot. The method includes: acquiring a training route corresponding to the user's current rehabilitation stage; setting multiple positioning markers in the rehabilitation area; establishing a coordinate system mapping between the physical area and digital space based on the positioning markers and the site parameters corresponding to the rehabilitation area to adjust the training route; collecting mixed EEG signals corresponding to the user; decoding and obtaining the movement intention information corresponding to the mixed EEG signals; collecting electromyography (EMG) signals of the user's affected limb and robot gait data corresponding to the lower limb rehabilitation robot; sensing depth information through a depth camera; fusing the training route, visual perception data, EMG signals of the affected limb, robot gait data, and movement intention information to generate a path adjustment command; and controlling the lower limb rehabilitation robot to autonomously execute walking movements according to the path adjustment command to complete rehabilitation training without human intervention.

Description

Technical Field

[0001] This application relates to the field of medical device technology, and in particular to a method and system for self-walking path recognition and planning for a lower limb rehabilitation robot. Background Technology

[0002] In the field of medical devices, autonomous walking technology for lower limb rehabilitation robots has long faced the dual challenges of "high dependence on human intervention" and "insufficient environmental adaptability." Currently, most rehabilitation robots adopt a "passive following of the patient's gait" or "remotely controlled path" model. The path planning logic is based solely on preset trajectories or single visual signals (such as depth cameras), failing to integrate the patient's physiological state and movement intentions in real time. For example, traditional methods only recognize simple commands (such as start / stop) through EEG signals, but lack quantitative analysis of the intensity of movement intentions (such as speed preferences and turning tendencies), leading to a disconnect between path planning and the patient's active rehabilitation needs. Solutions relying solely on preset routes struggle to cope with obstacles or boundary deviations, while purely visual perception solutions are susceptible to environmental interference such as lighting and occlusion, posing safety hazards. Furthermore, existing technologies do not dynamically bind the patient's rehabilitation stage (such as early, middle, and late stages) to path parameters. General-purpose path planning cannot meet the targeted training needs of different rehabilitation cycles, and requires continuous human supervision to prevent collision risks, limiting its widespread adoption in home settings.

[0003] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0004] This application provides a method and system for autonomous walking path recognition and planning for lower limb rehabilitation robots, aiming to solve the dual challenges of "high dependence on human intervention" and "insufficient environmental adaptability" that have long faced by autonomous walking technology of lower limb rehabilitation robots in the field of medical devices.

[0005] In a first aspect, embodiments of this application provide a method for self-walking path recognition and planning for a lower limb rehabilitation robot, the method comprising: Obtain the training route corresponding to the user's current rehabilitation stage, set multiple positioning markers in the rehabilitation site, and establish a coordinate system mapping between the physical site and the digital space based on the positioning markers and the site parameters corresponding to the rehabilitation site, so as to adjust the training route. Collect mixed EEG signals corresponding to the user, decode and obtain the movement intention information corresponding to the mixed EEG signals; collect electromyography signals of the affected limb corresponding to the user and robot gait data corresponding to the lower limb rehabilitation robot; perceive depth information through a depth camera, the depth information including at least site obstacle information, path deviation information and boundary information; By integrating the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and movement intention information, a path adjustment command is generated; the lower limb rehabilitation robot is controlled to autonomously execute walking movements according to the path adjustment command, thus completing rehabilitation training without human intervention.

[0006] In some embodiments, establishing a coordinate system mapping between the physical site and the digital space based on the location markers and the site parameters corresponding to the rehabilitation site includes: acquiring image feature information of the location markers through a depth camera, identifying the corner points or contour features of the location markers using an image processing algorithm, and, in conjunction with the pre-input site length, width, and height parameters, establishing a one-to-one mapping relationship between each location point of the physical site and the digital space coordinate system based on the principle of perspective transformation and coordinate transformation algorithm, and generating a site coordinate mapping table.

[0007] In some embodiments, adjusting the training route includes: matching the coordinates of key nodes of the preset training route with the actual site coordinates in the coordinate system mapping table, calculating the positional and angular deviations of the route nodes, interpolating and correcting the deviation nodes using a path smoothing algorithm, or adjusting the curvature of the route based on a Bezier curve fitting algorithm, so that the training route conforms to the spatial layout of the actual site.

[0008] In some embodiments, the decoding to obtain the motion intention information corresponding to the mixed EEG signal includes: sequentially performing bandpass filtering on the acquired mixed EEG signal to remove noise, extracting the energy spectrum features of the characteristic frequency band through wavelet transform or principal component analysis, and using support vector machine or convolutional neural network classification algorithms to identify the type of motion intention and quantify the level of intention intensity.

[0009] In some embodiments, the acquisition of electromyographic signals of the affected limb corresponding to the user and robot gait data corresponding to the lower limb rehabilitation robot includes: acquiring electromyographic signals through surface electrodes attached to the muscle groups of the affected limb, and extracting the root mean square value or average absolute value as muscle activity intensity features after amplification, rectification and normalization; acquiring gait data through inertial sensors installed on the robot joints, extracting stride length, stride frequency and joint angle change rate, and generating gait feature vectors.

[0010] In some embodiments, the perception of depth information via a depth camera includes: acquiring three-dimensional point cloud data of the site using the structured light or time-of-flight principle of the depth camera; removing outliers using voxel filtering; segmenting obstacle point clouds using Euclidean clustering algorithm; calculating the position coordinates, size, and distance of obstacles from the robot; and identifying the color or texture features of path boundaries using an edge detection algorithm to determine the direction and distance of path deviation.

[0011] In some embodiments, the process of fusing the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and motion intention information to generate path adjustment instructions includes: establishing a multi-source data fusion framework; performing time-series alignment and weight allocation on the training route coordinates, obstacle positions, electromyographic intensity, gait characteristics, and motion intention levels using a Kalman filter algorithm; generating an obstacle avoidance path based on a preset obstacle avoidance algorithm; dynamically adjusting the target speed in conjunction with the motion intention intensity; and generating motor speed and steering angle instructions.

[0012] In some embodiments, controlling the lower limb rehabilitation robot to autonomously execute walking movements according to path adjustment instructions to complete rehabilitation training without human intervention includes: converting the speed and steering parameters in the path adjustment instructions into differential speed control signals for the front wheel hub motors; driving the motors of the lower limb rehabilitation robot to perform steering and speed adjustment movements through pulse width modulation technology; adaptively adjusting the support direction of the rear omnidirectional wheels of the lower limb rehabilitation robot; monitoring feedback data from depth cameras and electroencephalogram (EEG) signals; and triggering braking or path replanning procedures when the distance to an obstacle is detected to be less than a safety threshold or the movement intention command is changed.

[0013] In some embodiments, obtaining the training route corresponding to the user's current rehabilitation stage includes: automatically identifying the rehabilitation stage based on the rehabilitation stage identifier input by the user or through historical training data, and retrieving the corresponding route template from a preset route database: the rehabilitation stage includes the initial, middle, and late stages; the initial stage matches a first speed range and a straight path template, the middle stage matches a second speed range and a turning path template, and the late stage matches a third speed range and a path template containing slopes or complex curves, and dynamically fine-tuning the route parameters based on the user's real-time physiological signals; wherein, the first speed range is smaller than the second speed range, and the second speed range is smaller than the third speed range.

[0014] Secondly, this application provides a self-walking path recognition and planning system for a lower limb rehabilitation robot, the system comprising: The route acquisition unit is used to acquire the training route corresponding to the current rehabilitation stage of the user, set multiple positioning markers in the rehabilitation site, and establish a coordinate system mapping between the physical site and the digital space based on the positioning markers and the site parameters corresponding to the rehabilitation site, so as to adjust the training route. The signal acquisition unit is used to acquire the user's corresponding mixed EEG signal, decode and obtain the movement intention information corresponding to the mixed EEG signal; acquire the user's corresponding affected limb electromyography signal and the robot gait data corresponding to the lower limb rehabilitation robot; and perceive depth information through a depth camera, wherein the depth information includes at least site obstacle information, path deviation information and boundary information. The training completion unit is used to integrate the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and movement intention information to generate path adjustment instructions; and control the lower limb rehabilitation robot to autonomously execute walking movements according to the path adjustment instructions to complete rehabilitation training without human intervention.

[0015] This application uses mixed EEG signals as the core, supplemented by EMG and gait multimodal physiological signals, to deeply bind the patient's active movement intentions such as starting, stopping, speed adjustment, and direction preference with path recognition and planning logic. It breaks through the design defect of existing technology that "only uses a single signal to control robot movements and has no autonomous path planning", so that the robot's autonomous walking is fully in line with the patient's rehabilitation ability and active intentions, and greatly improves the pertinence and compliance of rehabilitation training.

[0016] Adopting a dual-support mode of "preset route import + real-time perception by depth camera", the preset route can be accurately adapted to different rehabilitation scenarios such as home use and rehabilitation institution passages. The depth camera can capture path deviations and obstacles in real time, solving the pain points that a single preset route cannot cope with dynamic scenarios and that single visual perception is easily affected by environmental interference. It achieves stable autonomous walking without the need for additional personnel intervention, filling the gap in the existing technology in the fields of dual redundancy adaptation and autonomous path adjustment.

[0017] By deeply integrating path planning logic with the patient's rehabilitation stage and the intensity of their movement intention, the system can dynamically adjust preset route parameters and planning strategies according to different rehabilitation needs in the early stage (slow speed, small stride, straight path), the middle stage (medium speed, adjustable stride, simple turns), and the later stage (normal speed, custom stride, complex path). At the same time, it can optimize walking parameters in real time by combining the quantitative value of movement intention, which solves the problem of existing general path planning not fitting the rehabilitation scenario and ensures that the autonomous walking process meets the rehabilitation training goals simultaneously.

[0018] The path recognition and planning module integrates safety logic specific to rehabilitation scenarios. After the depth camera identifies the boundaries and obstacles of the rehabilitation scenario, it can combine multimodal physiological signals to judge the patient's coping ability and automatically trigger deceleration, stopping, or path detour. At the same time, it links with the original EEG dual detection mechanism. When the signal is invalid, it immediately stops walking, breaking through the limitation of existing robots that "require human supervision to prevent risks". This ensures the safety of autonomous walking and reduces the labor costs of rehabilitation institutions and the care pressure in home scenarios.

[0019] By reusing the original all-silicone flexible EEG acquisition, dual detection, and hybrid paradigm decoding modules, and only adding path recognition, planning, and depth camera vision modules, without modifying the original hardware structure, the problem of existing autonomous walking robots "requiring dedicated hardware and being incompatible with rehabilitation training systems" is solved. It has the advantages of high hardware reuse rate and low integration cost, and is easy to promote and apply on the basis of existing rehabilitation training systems.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic flowchart illustrating the steps of a self-walking path recognition and planning method for a lower limb rehabilitation robot provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a lower limb rehabilitation robot provided in one embodiment of this application; Figure 3 This is a schematic diagram of an elliptical trajectory provided in an embodiment of this application; Figure 4 This is a schematic diagram of a circular trajectory provided in one embodiment of this application; Figure 5 This is a schematic diagram of a square trajectory provided in an embodiment of this application; Figure 6 This is a schematic diagram of a linear reciprocating trajectory provided in an embodiment of this application; Figure 7 This is a schematic diagram of an 8-shaped trajectory provided in one embodiment of this application; Figure 8 This is a schematic diagram of the placement of positioning markers for an elliptical trajectory provided in one embodiment of this application; Figure 9 This is a schematic diagram of the placement of positioning markers for a circular trajectory provided in one embodiment of this application; Figure 10 This is a schematic diagram of the placement of positioning markers for a square trajectory provided in one embodiment of this application; Figure 11 This is a schematic diagram of the placement of positioning markers for a linear reciprocating trajectory provided in one embodiment of this application; Figure 12 This is a schematic diagram of the placement of positioning markers for a figure-eight trajectory provided in one embodiment of this application; Figure 13 This is a schematic block diagram of a self-walking path recognition and planning system for a lower limb rehabilitation robot provided in one embodiment of this application; Figure 14 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0026] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0027] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0028] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] In the field of medical devices, autonomous walking technology for lower limb rehabilitation robots has long faced the dual challenges of "high dependence on human intervention" and "insufficient environmental adaptability." Currently, most rehabilitation robots adopt a "passive following of the patient's gait" or "remotely controlled path" model. The path planning logic is based solely on preset trajectories or single visual signals (such as depth cameras), failing to integrate the patient's physiological state and movement intentions in real time. For example, traditional methods only recognize simple commands (such as start / stop) through EEG signals, but lack quantitative analysis of the intensity of movement intentions (such as speed preferences and turning tendencies), leading to a disconnect between path planning and the patient's active rehabilitation needs. Solutions relying solely on preset routes struggle to cope with obstacles or boundary deviations, while purely visual perception solutions are susceptible to environmental interference such as lighting and occlusion, posing safety hazards. Furthermore, existing technologies do not dynamically bind the patient's rehabilitation stage (such as early, middle, and late stages) to path parameters. General-purpose path planning cannot meet the targeted training needs of different rehabilitation cycles, and requires continuous human supervision to prevent collision risks, limiting its widespread adoption in home settings.

[0030] Therefore, a method is urgently needed to solve at least one of the above problems.

[0031] To solve the above problem, please refer to Figure 1 This application provides a method for self-walking path recognition and planning for lower limb rehabilitation robots, applied to computer equipment. The computer equipment can be deployed on a single server or server cluster. It can also be deployed on handheld terminals, laptops, wearable devices, or robots, etc. It should be noted that all information involved in the method provided in this application is extracted with the authorization of the relevant users and in accordance with relevant regulations, and will not infringe on user privacy.

[0032] Meanwhile, the provided method is used to train such as Figure 2 The lower limb rehabilitation robot shown in this application does not limit the type of lower limb rehabilitation robot.

[0033] The provided method for self-walking path recognition and planning for lower limb rehabilitation robots includes steps S101 to S103. Details are as follows: Step S101. Obtain the training route corresponding to the current rehabilitation stage of the user, set multiple positioning markers in the rehabilitation site, and establish a coordinate system mapping between the physical site and the digital space based on the positioning markers and the site parameters corresponding to the rehabilitation site, so as to adjust the training route.

[0034] Specifically, this step is the preliminary benchmark establishment stage for the autonomous walking of the lower limb rehabilitation robot, which achieves two core functions: first, to complete the pre-configuration of a training route that is accurately adapted to the user's current rehabilitation stage and establish the target benchmark for rehabilitation training; second, to complete the high-precision coordinate system mapping between the physical site and the digital space through positioning markers, thereby establishing a unified spatial benchmark for subsequent path recognition, deviation comparison, and dynamic planning, and solving the problem from the source that the existing technology's preset routes are out of sync with the patient's rehabilitation needs and that the lack of a unified spatial positioning benchmark leads to insufficient path planning accuracy.

[0035] This step is divided into three sub-steps: rehabilitation training route matching and pre-configuration, site positioning marker placement, and coordinate system mapping establishment and route adaptation adjustment. 1. Acquisition and pre-configuration of training routes matched to the rehabilitation phase: (1) Rehabilitation stage classification and route matching: The system pre-stores the rehabilitation stage classification standard based on the Brunnstrom classification of hemiplegia, which is divided into three core rehabilitation cycles: initial stage (stages 1-3), intermediate stage (stages 4-5), and late stage (stage 6). Each cycle corresponds to a dedicated parametric training route library. Operators input the user's current rehabilitation assessment results through the host computer, and the system automatically matches the corresponding training route set. At the same time, it supports manual customization of routes to adapt to personalized rehabilitation goals.

[0036] (2) Parametric definition and storage of training routes: Each training route is an editable digital trajectory. The core parameters include trajectory type (straight line, fixed / variable radius turn, slope adaptation, obstacle avoidance simulation), total trajectory length, baseline walking speed, baseline stride, turning angle, turning radius, start and stop nodes, slope adaptation coefficient, etc. The route data is pre-stored in the non-volatile memory of the preset route storage and retrieval module, which can be associated with the user's personal rehabilitation records and supports offline retrieval, historical reuse and online iterative updates.

[0037] (3) Baseline route import: After the user's rehabilitation stage matching is completed, the system automatically imports the corresponding training route into the calculation unit of the path recognition and planning module and locks it as the baseline walking trajectory for this rehabilitation training.

[0038] 2. Standardized layout of location markers in rehabilitation areas: (1) Positioning mark selection: High contrast, anti-light interference circular reflective plane mark with a diameter of 80mm-150mm. The reflective layer uses engineering grade reflective film, and the bottom layer is an adhesive flexible substrate that can be adapted to various floor materials such as ceramic tiles, wood flooring, and plastic flooring.

[0039] (2) Marking arrangement rules: In conventional rehabilitation scenarios, a 5-point positioning reference array of "four corners + center" is adopted, that is, one mark is pasted at each of the four corners of the rehabilitation site and one mark is pasted at the geometric center of the site; for long strip rehabilitation channels with a length of more than 10m, one mark is added at the midpoint of each long side of the channel to form a 7-point positioning array to improve the accuracy of long-distance trajectory calibration; all mark pasting planes are flush with the walking ground, without protrusions or obstructions, and the error of the center point of the mark does not exceed ±5mm.

[0040] (3) Input of basic parameters of the site: The operator inputs the measured parameters of the rehabilitation site into the host computer, including the site length / width, boundary coordinates, passage width, slope parameters, fixed obstacle positions, etc., as auxiliary references for coordinate system calibration.

[0041] 3. Establishment of physical site-digital coordinate system mapping and route adjustment: (1) Identification feature acquisition and extraction: The system controls the RGB-D depth camera (depth detection range 0.1m-10m, frame rate ≥30fps, image resolution ≥1920×1080) mounted on the front of the robot body to sequentially acquire RGB images and depth point cloud data of all positioning identifiers; through Gaussian filtering, binarization and contour extraction image preprocessing algorithms, the center pixel coordinates of each identifier are extracted, and the three-dimensional spatial coordinates of the identifier relative to the robot are obtained through depth ranging.

[0042] (2) Coordinate system mapping and calibration: The world coordinate system origin (0,0,0) is set with the lower left corner of the rehabilitation site as the location marker, the long side of the site as the X-axis, the short side as the Y-axis, and the vertical upward direction as the Z-axis. The physical site world coordinate system is established; the camera coordinate system is set with the optical center of the depth camera as the origin. The camera intrinsic and extrinsic parameters are calibrated by Zhang Zhengyou calibration method to eliminate lens distortion; based on the extracted location marker world coordinates and camera coordinates, the transformation matrix of the two coordinate systems is solved by the PNP algorithm to complete the coordinate system mapping between the physical site and the digital space, and a digital twin virtual rehabilitation site is established. The spatial positioning error after calibration does not exceed ±10mm.

[0043] (3) Training route adaptation and adjustment: The system maps the pre-imported benchmark training route to the digital twin virtual field, and automatically adapts and adjusts the route based on the actual measured parameters of the field and the actual location of the positioning markers. This includes adjusting the total length of the straight trajectory according to the actual length of the field, adjusting the turning radius according to the field boundary, and adding avoidance nodes according to the location of fixed obstacles. This ensures that the adjusted route is fully adapted to the current physical field and does not deviate from the user's training goals in the rehabilitation stage. The adjusted route is then locked as the benchmark trajectory for this training.

[0044] Step S102. Collect the user's corresponding mixed EEG signal, decode to obtain the movement intention information corresponding to the mixed EEG signal; collect the user's corresponding affected limb electromyography signal and the robot gait data corresponding to the lower limb rehabilitation robot; perceive depth information through a depth camera, the depth information including at least site obstacle information, path deviation information and boundary information.

[0045] Specifically, this step is the core of multi-source perception for autonomous walking control, achieving three main functions: First, it completes high-precision acquisition of hybrid EEG signals and full-dimensional decoding of motor intentions to obtain the core control basis for autonomous walking; second, it completes synchronous acquisition of electromyography signals of the affected limb and robot gait data to obtain the physiological and state basis for walking posture correction; and third, it completes real-time visual depth perception of the rehabilitation scene to obtain environmental basis for environmental safety and path deviation. At the same time, it establishes a strict time synchronization mechanism for multi-source signals to solve the pain points of existing technologies such as single dimension of motor intention decoding, asynchronous multimodal signals, and incomplete environmental perception dimensions.

[0046] This step is divided into four main acquisition and decoding units and a unified time synchronization mechanism, specifically including: 1. Hybrid Paradigm EEG Signal Acquisition and Motor Intent Decoding: (1) Hybrid paradigm EEG signal acquisition: A flexible silicone EEG acquisition cap was used, and the acquisition electrodes were arranged according to the international 10-20 lead system, focusing on the core leads Fz, Cz, Pz, and Oz. The acquisition frequency band was 0.5Hz-50Hz, and the sampling rate was ≥500Hz. The acquired hybrid paradigm EEG signals were fused with the motor imagery (MI) paradigm and the P300 evoked potential paradigm. The motor imagery paradigm was used to decode continuous motor intentions, and the P300 paradigm was used to decode discrete control commands, achieving dual-dimensional coverage of "continuous intention quantification + discrete command recognition".

[0047] (2) Preprocessing of dual detection of EEG signals: The raw EEG signals are transmitted to the dual detection module of EEG signals and two-level preprocessing is performed: The first level is noise filtering, which removes noise such as power frequency, electrooculogram, electromyogram, and motion interference by adaptive wavelet threshold filtering, and separates invalid artifacts by independent component analysis (ICA); The second level is feature extraction, which extracts EEG features of the motor imagery paradigm by common spatial pattern (CSP), and extracts feature potentials of the P300 paradigm by superposition averaging algorithm to obtain an effective EEG feature set.

[0048] (3) Decoding output of motor intention information: The EEG feature set is input into the pre-trained CNN-LSTM hybrid decoding model, and outputs two types of core motor intention information. The decoding results are updated in real time with a period of 100ms: ① Discrete walking control commands, including 7 types of core commands such as start, stop, acceleration, deceleration, left turn, right turn, and emergency braking, with a recognition accuracy of ≥95%; ② Continuous motor intention intensity quantification value, including walking speed preference coefficient (0-1, corresponding to 0%-120% of the baseline speed), turning tendency coefficient (-1 to +1, corresponding to the maximum left / right turning angle), and exercise tolerance coefficient (0-1, corresponding to the user's exercise fatigue level).

[0049] 2. Synchronous acquisition of electromyography (EMG) signals of the affected limb: A multi-channel surface EMG acquisition module with a sampling rate ≥1000Hz is used. The acquisition electrodes are attached to the surface of four core muscle groups of the user's affected limb: quadriceps femoris, hamstrings, tibialis anterior, and gastrocnemius. The acquisition process is strictly synchronized with the EEG signal. After filtering, rectification, and smoothing preprocessing, the root mean square (RMS) and integrated EMG (iEMG) features are extracted to quantify the degree of muscle activation, muscle fatigue, and motor coordination of the affected limb. The data are transmitted in real time to the path recognition and planning module as the basis for correcting walking posture and stride.

[0050] 3. Synchronous acquisition of robot gait data: Gait acquisition units are integrated into the robot's walking mechanism, plantar pressure sensors, and motor encoders, including hub motor encoders, plantar pressure sensors, and a six-axis IMU inertial measurement unit. The acquisition unit synchronously acquires real-time gait data of the robot at a sampling rate of 200Hz, including real-time walking speed, stride length, stride frequency, turning angle, center of gravity offset, plantar pressure distribution, and robot tilt angle. Through feature extraction algorithms, the deviation between the current gait and the reference gait, user center of gravity stability, and walking smoothness data are identified and transmitted in real time to the path recognition and planning module as an auxiliary basis for path adjustment and posture correction.

[0051] 4. Real-time perception of visual depth information by depth camera: (1) Real-time acquisition: The RGB-D depth camera at the front end of the robot acquires RGB images and depth point cloud data of the 0.1m-8m front and 60° fan-shaped areas to the left and right in real time at a frame rate of 30fps. There are no blind spots. The acquisition process is synchronized with other multi-source signals in time, and the synchronization error is ≤20ms.

[0052] (2) Depth information extraction and recognition: The system processes the collected images and point cloud data in real time and extracts three types of core depth information, which are updated in 30ms cycles: ① Site obstacle information: By combining YOLOv8 target detection with point cloud segmentation algorithm, dynamic / static obstacles are identified, and the three-dimensional coordinates, contour size, relative distance and movement trend of the obstacles are output, and the safe avoidance distance is calculated; ② Path deviation information: By visual feature matching, the real-time ground trajectory is compared with the digital space baseline route, and the lateral / longitudinal deviation and heading angle deviation are output; ③ Boundary information: By semantic segmentation algorithm, the site boundary line, wall and passage edge are identified, and the relative distance between the robot and the boundary is output to judge the risk of crossing the boundary.

[0053] 5. Multi-source signal time synchronization mechanism: All signal acquisition is triggered by the unified hardware clock of the robot's main controller. A unified timestamp is added to the EEG, EMG, gait, and visual data to ensure that the time synchronization error of multi-source data is ≤20ms, avoiding path planning delays and control deviations caused by signal asynchrony.

[0054] Step S103. Integrate the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and movement intention information to generate a path adjustment command; control the lower limb rehabilitation robot to autonomously execute walking movements according to the path adjustment command to complete rehabilitation training without human intervention.

[0055] Specifically, this step is the core control closed-loop link of the present invention, which realizes three main functions: First, it completes the priority fusion of multi-source data through an adaptive weighted fusion algorithm to establish a decision-making system that takes into account rehabilitation goals, user intentions, environmental safety, and limb status; second, it completes dynamic path planning specific to the rehabilitation scenario and generates precise path adjustment instructions; third, it realizes the full closed-loop control of the robot's autonomous walking, while integrating multiple redundant safety protection mechanisms, and finally constructs a full closed-loop logic of "intention-driven - environmental perception - state correction - path planning - execution feedback" to achieve autonomous walking without human intervention, solving the core pain points of existing technologies such as path planning being disconnected from the patient's physiological state, insufficient dynamic adjustment ability, and high dependence on human intervention.

[0056] This step is divided into four sub-stages: multi-source data adaptive fusion, dynamic path planning and command generation, robot autonomous walking closed-loop control, and end-to-end safety protection. Specifically, it includes: 1. Adaptive Weighted Fusion of Multi-Source Data: After receiving multi-source data, the path identification and planning module processes the data using an adaptive weighted fusion algorithm to provide the optimal decision-making basis for path planning. (1) Basic weight classification: The basic weight is preset according to the core control priority of the data. Among them, the motor intention information (EEG decoding result) is the core control source with a basic weight of 40%; the benchmark training route is the planning benchmark source with a basic weight of 25%; the visual perception data is the safety environment source with a basic weight of 20%; and the electromyography and gait data of the affected limb are the posture correction source with a combined basic weight of 15%.

[0057] (2) Dynamic adaptive adjustment of weights: The system dynamically adjusts the weights according to the real-time scenario and user status: ① When visual perception detects obstacles or risks of crossing boundaries, the weight of visual data is increased to 35%, and the weight of the baseline route is reduced to 10%, prioritizing safety; ② When electromyography signals detect decreased muscle strength and muscle fatigue in the affected limb, the weight of electromyography + gait data is increased to 25%, and the weight of speed preference coefficient is reduced simultaneously to limit the maximum walking speed; ③ When EEG decodes emergency braking or stop commands, the weight of motor intention information is increased to 100%, prioritizing the response to user active commands.

[0058] (3) Fusion result output: After fusion, the baseline trajectory parameters, environmental constraint parameters, user status parameters, and control priority parameters are output as the input basis for dynamic path planning.

[0059] 2. Dynamic Path Planning and Path Adjustment Command Generation: The path identification and planning module uses a 100ms planning cycle to perform real-time dynamic path planning based on fused data and generate standardized path adjustment commands. The core process consists of four stages: (1) Baseline path matching and deviation correction: Compare the robot's real-time position with the baseline training route, calculate the path's lateral and longitudinal deviations and heading angle deviations, generate the path deviation correction amount based on the PID control algorithm, adjust the robot's heading angle and walking direction to ensure that the trajectory fits the baseline route, and the maximum deviation correction threshold does not exceed ±150mm to avoid deviating from the rehabilitation training goal.

[0060] (2) Obstacle avoidance and boundary constraint planning: Safe path planning is performed based on visual perception data: ① Static obstacles: When the relative distance is less than the preset safe distance (preferably 2m), a collision-free avoidance path is generated based on the artificial potential field method, and the walking direction and turning radius are adjusted. During the avoidance process, the walking speed does not exceed 50% of the base speed; ② Dynamic obstacles: The motion trajectory is predicted based on the Kalman filter algorithm, and deceleration / avoidance instructions are generated in advance. When the predicted collision risk exceeds the threshold, the speed is immediately reduced to a stop; ③ Scene boundary: When the relative distance with the boundary is less than the safe threshold (preferably 500mm), a heading correction instruction is immediately generated to limit the turning range and avoid crossing the boundary.

[0061] (3) Adjustment of walking parameters for motor intention adaptation: The walking parameters are dynamically adjusted based on the motor intention information decoded by EEG: ① Discrete control instructions: When the start, stop, turn and other instructions are received, the corresponding adjustment instructions are generated immediately with a response delay of ≤200ms; ② Continuous intention quantification value: The walking speed is adjusted in real time based on the speed preference coefficient (the adjustment range is 30%-120% of the base speed), the turning angle and turning radius are adjusted based on the turning tendency coefficient, and the speed is automatically reduced, the stride is shortened or a pause is triggered when the user's fatigue exceeds the threshold based on the exercise tolerance coefficient.

[0062] (4) Posture correction and adjustment for the condition of the affected limb: micro-correction of walking posture based on electromyography and gait data: ① Based on the degree of muscle activation of the affected limb, dynamically adjust the stride and the duration of the support phase. When the muscle strength is insufficient, extend the support phase and shorten the stride to reduce the weight-bearing of the affected limb; ② Based on gait deviation and center of gravity shift data, fine-tune the walking speed on the left and right sides and the support strength of the foot to correct abnormal gait; ③ When abnormal gait or instability of the center of gravity is detected, immediately generate a deceleration to stop command.

[0063] (5) Final output of path adjustment instructions: Integrate the above correction, avoidance, adaptation and correction results to generate a standardized path adjustment instruction set, including walking speed, turning angle, stride adjustment, support / swing phase duration, start and stop control instructions. All instructions are digital signals that can be directly recognized by the robot execution terminal and are synchronously output to the robot main controller.

[0064] 3. Closed-loop control of autonomous walking of lower limb rehabilitation robot: (1) Command decomposition and drive execution: After receiving the path adjustment command, the robot main controller decomposes it into control signals for the corresponding actuators: In terms of walking drive, the left and right hub motors of the front wheel output speed and angle signals, and speed adjustment and steering are achieved through differential control of the left and right motors. The rear universal wheel follows and adapts to the steering; In terms of gait assistance, the hip and knee joint drive motors output stride and support phase duration adjustment signals to adapt to the movement ability of the affected limb.

[0065] (2) Execution feedback and closed-loop correction: During the robot's walking process, the gait acquisition unit collects the actual walking status data (actual speed, turning angle, stride, position) in real time and sends it back to the path recognition and planning module; the system compares the actual execution status with the planning target, identifies the execution deviation, and completes the instruction closed-loop correction in the next planning cycle, forming a full closed-loop control of "planning-execution-feedback-correction".

[0066] (3) Autonomous operation without human intervention throughout the entire rehabilitation training process: The robot autonomously completes path planning, adjustment, avoidance, start and stop actions based on multi-source fusion data, without the need for manual operation or remote intervention; it only stops walking and triggers an alarm when an emergency safety risk, abnormal signal, or user triggers emergency braking is detected, and the rest of the process is executed autonomously.

[0067] 4. Dedicated end-to-end security protection for rehabilitation scenarios: (1) Graded safety threshold control: Graded safety thresholds are preset for users at different rehabilitation stages, including maximum walking speed, minimum avoidance distance, maximum turning angle, and boundary warning distance. Stricter safety thresholds are adopted for users in the early stage of rehabilitation to further reduce risks.

[0068] (2) Multi-redundant emergency braking mechanism: Three-level braking trigger conditions are set, which are redundant to each other: ① Level 1 trigger: When the EEG decodes the emergency braking or stop command, the braking is triggered immediately; ② Level 2 trigger: When the visual perception detects an unavoidable obstacle or risk of crossing the boundary, the braking is triggered immediately; ③ Level 3 trigger: When the electromyography and gait data detect a sudden drop in muscle strength and instability of the center of gravity, the braking is triggered immediately; the braking response time is ≤100ms.

[0069] (3) Signal abnormality protection mechanism: When the EEG signal is detected to be invalid for three consecutive decoding cycles, the electromyography electrode is detached, or the depth camera has no effective image, the robot is immediately controlled to decelerate smoothly to a stop, and an alarm message is sent to the host computer to avoid the risk of loss of control caused by signal abnormality.

[0070] In some embodiments, establishing a coordinate system mapping between the physical site and the digital space based on the location markers and the site parameters corresponding to the rehabilitation site includes: acquiring image feature information of the location markers through a depth camera, identifying the corner points or contour features of the location markers using an image processing algorithm, and, in conjunction with the pre-input site length, width, and height parameters, establishing a one-to-one mapping relationship between each location point of the physical site and the digital space coordinate system based on the principle of perspective transformation and coordinate transformation algorithm, and generating a site coordinate mapping table.

[0071] This embodiment is the core implementation scheme for mapping the physical site to the digital spatial coordinate system in step S101. It addresses the pain points of existing rehabilitation robot site calibration processes, such as complex procedures, low coordinate transformation accuracy, and large single-view perspective distortion errors. By combining depth camera visual feature extraction, perspective transformation principles, and coordinate transformation algorithms, it achieves rapid establishment of a site coordinate system without a dedicated calibration plate, completing a one-to-one correspondence between physical site positioning points and digital spatial coordinates. Finally, it generates a site coordinate mapping table that can be called in real time, providing a unified and high-precision spatial reference for subsequent path planning, deviation comparison, and trajectory control. After calibration, the coordinate transformation error can be controlled within ±10mm.

[0072] The system controls an RGB-D depth camera mounted on the front of the lower limb rehabilitation robot to acquire continuous frame images covering all site positioning markers at a frame rate of ≥30fps and an image resolution of ≥1920×1080, and simultaneously acquire the corresponding three-dimensional depth point cloud data. During the acquisition process, the robot can traverse the edge of the site at a low speed to ensure that all positioning markers are acquired without obstruction or severe perspective distortion, thus avoiding feature recognition errors caused by single-view acquisition.

[0073] The acquired images are sequentially processed using Gaussian blur filtering (to remove lighting noise and lens noise), grayscale processing, and adaptive threshold binarization to separate the positioning marker from the ground background. A contour search algorithm is used to extract the closed contour of the positioning marker, and valid contours that meet the preset size and shape characteristics are selected, while invalid contours caused by environmental interference are eliminated. For square positioning markers, the Shi-Tomasi algorithm is used to extract the coordinates of the four corner points, and for circular positioning markers, the Hough circle transform is used to extract the coordinates of the center of the circle, which serve as the core feature points of the positioning marker. At the same time, the three-dimensional spatial depth value corresponding to each feature point is obtained through depth point cloud data.

[0074] The physical site's world coordinate system is defined by taking the lower left corner of the site as the origin O(0,0,0), the longer side of the site as the X-axis, the shorter side as the Y-axis, and the vertical upward direction as the Z-axis. Combined with the pre-inputted length, width, and height parameters of the site, the coordinate system boundaries and passable range of the physical site are determined. The digital spatial coordinate system is defined by adopting the Cartesian coordinate system, which is of the same origin as the physical world coordinate system. The unit scale is completely consistent with the physical site, ensuring the scale uniformity of spatial mapping.

[0075] Based on the principle of perspective transformation, the positioning markers in the image pixel coordinate system are converted into actual coordinates in the physical world coordinate system. By solving the homography matrix, the basic transformation relationship between pixel coordinates and world coordinates is established. Combined with the PNP algorithm, the camera extrinsic parameters (rotation matrix, translation vector) are solved to complete the accurate transformation from the camera coordinate system to the world coordinate system. For feature points acquired from multiple perspectives, the transformation matrix is ​​optimized by the bundle adjustment method to further reduce the transformation error caused by perspective distortion.

[0076] The system maps all the physical world coordinates, digital space coordinates, pixel coordinates, coordinate transformation matrices, and error correction coefficients of the positioning markers to generate a structured site coordinate mapping table. The mapping table is stored in the system memory and can be called in real time to quickly convert any physical site location to a digital space location, providing a unified spatial transformation benchmark for subsequent path planning.

[0077] In some embodiments, adjusting the training route includes: matching the coordinates of key nodes of the preset training route with the actual site coordinates in the coordinate system mapping table, calculating the positional and angular deviations of the route nodes, interpolating and correcting the deviation nodes using a path smoothing algorithm, or adjusting the curvature of the route based on a Bezier curve fitting algorithm, so that the training route conforms to the spatial layout of the actual site.

[0078] This embodiment is the core implementation scheme for the site adaptation adjustment of the preset training route in step S101. It addresses the pain points of mismatch between the preset route and the actual site space, abrupt route turning points, and failure to meet the safety requirements for rehabilitation walking. Through the entire process of key node coordinate matching, deviation quantification calculation, path smoothing interpolation correction, and Bézier curve curvature fitting, it achieves accurate adaptation of the preset training route to the actual site. This ensures that the adjusted route fully conforms to the rehabilitation training goals, adapts to the actual site boundaries, eliminates the risk of collision, and guarantees the smoothness of the walking trajectory, avoiding sudden turns and stops that could cause the patient to lose their center of gravity.

[0079] The key nodes of the preset training route are extracted by retrieving the preset training route matching the user's rehabilitation stage from the preset route storage and retrieval module. The core key nodes of the route are extracted, including the start point, end point, turning point, start and stop point, slope change point, and speed adjustment point. Each key node contains preset three-dimensional coordinates, heading angle, speed threshold, and stride parameters. At the same time, the preset coordinates of the key nodes are converted into theoretical coordinates in the digital space coordinate system through the field coordinate mapping table generated in Example 1.

[0080] Coordinate matching and deviation quantification calculation involves matching and verifying the transformed theoretical coordinates of key nodes with the actual site boundary coordinates, fixed obstacle coordinates, and positioning marker coordinates in the site coordinate mapping table. The position deviation (Euclidean distance between the theoretical coordinates and the actual passable area coordinates) and angle deviation (the angle between the preset heading angle at the key node and the actual passable direction of the site) of each key node are calculated sequentially. At the same time, it is verified whether the key nodes exceed the site boundary or spatially overlap with fixed obstacles. Nodes with position deviation > 150mm and angle deviation > 15° are marked as nodes to be corrected, and nodes that cross the boundary / overlap are marked as invalid nodes.

[0081] Interpolation smoothing correction of deviation nodes is performed by using a cubic spline interpolation algorithm for nodes whose position and angle deviations exceed the threshold. First, based on the actual passable area of ​​the site, the target correction coordinates of the node to be corrected are determined to ensure that the corrected node is within the passable area and its trajectory is continuous with the preceding and following nodes. Then, through cubic spline interpolation, several transition interpolation points are generated between adjacent key nodes to fill the trajectory gaps between nodes, eliminate abrupt changes in trajectory inflection points, and ensure the continuity of the route in terms of position, heading angle, and curvature. The position deviation of the corrected node is controlled within ±50mm, and the angle deviation is controlled within ±5°.

[0082] The route curvature is adjusted by Bézier curve fitting. This is done on training routes including turns and curves, as well as route segments after invalid node replacements. The algorithm uses the start, end, and turning control points of the route as anchor points for the Bézier curve. The coordinates of the control points are adjusted based on the actual site's passage width and turning space to generate second- or third-order Bézier curves, resulting in a smooth turning trajectory. By adjusting the order of the Bézier curve and the position of the control points, the route curvature is dynamically adjusted to ensure that the turning radius meets the user's rehabilitation stage requirements (initial turning radius ≥2m, mid-stage ≥1.5m, late-stage ≥1m), while avoiding the trajectory exceeding the site boundaries and colliding with obstacles. The fitted route curvature changes continuously without abrupt changes.

[0083] After the correction and fitting are completed, the adjusted route is verified and locked. The entire trajectory is verified to ensure that the route is within the passable area of ​​the site, without overlapping obstacles, and that the curvature and speed parameters match the user's rehabilitation stage. After the verification is passed, the key nodes, trajectory coordinates and motion parameters of the adjusted route are updated to the path recognition and planning module and locked as the benchmark walking route for this rehabilitation training.

[0084] In some embodiments, the decoding to obtain the motion intention information corresponding to the mixed EEG signal includes: sequentially performing bandpass filtering on the acquired mixed EEG signal to remove noise, extracting the energy spectrum features of the characteristic frequency band through wavelet transform or principal component analysis, and using support vector machine or convolutional neural network classification algorithms to identify the type of motion intention and quantify the level of intention intensity.

[0085] This embodiment is the core implementation scheme for decoding hybrid EEG signals in step S102. It addresses the shortcomings of existing EEG decoding methods, which can only recognize simple discrete commands, cannot quantify intent intensity, and have poor anti-interference capabilities. Through a three-level processing flow of "bandpass filtering denoising - feature frequency band extraction - classification recognition + intensity quantization," it achieves full-dimensional decoding of hybrid EEG signals, simultaneously outputting discrete motion intent types and continuous intent intensity levels. This provides precise active control basis for robot autonomous walking, with a single-frame decoding latency ≤50ms and a classification recognition accuracy ≥95%.

[0086] Bandpass filtering denoising of mixed EEG signals involves preprocessing the raw mixed EEG signals (sampling rate ≥ 500 Hz, acquisition frequency band 0.5 Hz-50 Hz) acquired in real time by the EEG acquisition module. An 8th-order Butterworth bandpass filter is used, with the passband frequency band set to 0.5 Hz-30 Hz, to filter out 50 Hz power frequency interference, high-frequency EMG artifacts, and low-frequency baseline drift noise. At the same time, an adaptive noise cancellation algorithm is used to remove artifact signals caused by EEG, ECG, and motion interference. The filtered signals are then subjected to baseline correction and normalization to normalize the signal amplitude to the [-1, 1] interval, eliminating the differences in signal amplitude between different leads and different acquisition time periods.

[0087] Feature extraction of energy spectrum in characteristic frequency bands involves extracting effective features from preprocessed EEG signals using wavelet transform or principal component analysis (PCA). For wavelet transform, the db4 wavelet basis function is used to perform a 5-level wavelet decomposition of the EEG signal, yielding wavelet coefficients for four core characteristic frequency bands: δ (0.5-4Hz), θ (4-8Hz), α (8-13Hz), and β (13-30Hz). The energy value and energy percentage of each frequency band's wavelet coefficients are calculated to construct an energy spectrum feature vector. For principal component analysis, a covariance matrix is ​​constructed for the multi-lead EEG signal, and eigenvalues ​​and eigenvectors are solved. Principal components with a cumulative contribution rate ≥90% are selected as feature vectors, reducing data dimensionality while retaining core motor intention-related features. For mixed-paradigm EEG signals, the β-band energy features of the motor imagery paradigm and the P300 peak potential features of the P300 paradigm are extracted separately and fused to obtain a multi-dimensional feature set.

[0088] Motion intent type classification and recognition employs pre-trained support vector machine (SVM) or convolutional neural network (CNN) classification algorithms to classify and identify the extracted feature set, outputting the motion intent type: The SVM classification method adopts a one-to-one multi-class SVM model with radial basis function (RBF) as the kernel function. During pre-training, the model parameters are optimized by labeled EEG sample data to achieve the recognition of 7 core motor intention types: start, stop, acceleration, deceleration, left turn, right turn, and emergency braking. The CNN classification method uses a lightweight one-dimensional CNN network, which includes two convolutional layers, two pooling layers, one fully connected layer and one softmax classification layer. It takes multi-lead EEG feature sequences as input and outputs the classification results of motion intention type end-to-end, reducing the computational load of the model and ensuring real-time performance.

[0089] While identifying the type of intent, the intensity of the motion intent is quantified and graded: based on the energy change amplitude of the characteristic frequency band, the matching degree between the feature value and the standard template, the intent confidence and intensity coefficient are calculated; the intensity of the motion intent is divided into 5 levels (level 1-5, the higher the level, the higher the intent intensity), and the intensity coefficient is normalized to the 0-1 range, corresponding to different control parameters: acceleration / deceleration intent: the intensity level corresponds to the speed adjustment range (level 1 corresponds to 10% of the base speed adjustment, level 5 corresponds to 50% of the base speed adjustment); left turn / right turn intent: the intensity level corresponds to the steering angle adjustment range (level 1 corresponds to 5° steering, level 5 corresponds to 30° steering); motion tolerance: the intensity level corresponds to the user's motion fatigue level, the lower the level, the higher the fatigue level, and automatic speed reduction adjustment is triggered.

[0090] The decoding results are output in real time with a period of 100ms to complete one complete EEG signal decoding. The motor intention type, intention intensity level, and intensity coefficient are output synchronously and transmitted to the path recognition and planning module in real time as the core basis for path planning and walking control.

[0091] In some embodiments, the acquisition of electromyographic signals of the affected limb corresponding to the user and robot gait data corresponding to the lower limb rehabilitation robot includes: acquiring electromyographic signals through surface electrodes attached to the muscle groups of the affected limb, and extracting the root mean square value or average absolute value as muscle activity intensity features after amplification, rectification and normalization; acquiring gait data through inertial sensors installed on the robot joints, extracting stride length, stride frequency and joint angle change rate, and generating gait feature vectors.

[0092] This embodiment is the core implementation scheme for collecting electromyographic signals of the affected limb and robot gait data in step S102. It addresses the shortcomings of existing technologies that rely solely on electroencephalogram (EEG) signals and lack feedback on limb status and robot gait, leading to a mismatch between walking posture and the patient's limb capabilities. By collecting electromyographic signals through surface electrodes and extracting muscle activity intensity features, and by collecting gait data through inertial sensors and encoders to generate standardized gait feature vectors, it provides quantitative data on limb status and robot operation status for subsequent path adjustment and posture correction, ensuring the adaptability and safety of rehabilitation training.

[0093] The acquisition and preprocessing of electromyography (EMG) signals of the affected limb were performed using a multi-channel surface EMG acquisition module with a sampling rate of ≥1000Hz. Disposable Ag / AgCl surface electrodes were used for acquisition and were attached to the muscle bellies of the four core motor muscle groups (quadriceps femoris, hamstrings, tibialis anterior, and gastrocnemius) of the affected limb according to their anatomical locations. Reference electrodes were attached to bony prominences of the affected limb (such as the tibial tuberosity). The electrode spacing was set to 20mm. Before attachment, the skin was cleaned and exfoliated to reduce contact impedance. The acquired raw EMG signals were first differentially amplified by an instrumentation amplifier (1000x amplification) to filter out common-mode interference, and then bandpass filtered (20Hz-500Hz) to remove power frequency and motion artifact noise, resulting in preprocessed EMG signals.

[0094] Electromyography (EMG) signal feature quantization extraction involves full-wave rectification of the preprocessed EMG signal to convert the bidirectional signal into a unidirectional positive signal, followed by low-pass filtering (cutoff frequency 10Hz) to obtain a smooth EMG envelope signal. Feature extraction is then performed on the envelope signal, focusing on two core types of quantization features: Root mean square (RMS): The root mean square value of the electromyographic signal is calculated in each time window of 200ms to quantify the real-time muscle activation and muscle strength of the corresponding muscle group. Mean Absolute Value (MAV): Calculates the average absolute value of electromyographic signals within each time window, representing the sustained intensity of muscle activity; Simultaneously, the median frequency (MF) of the electromyographic signal is extracted to determine muscle fatigue. A continuous decrease in the median frequency indicates muscle fatigue. The extracted features are used to generate an electromyographic feature vector in time sequence, which is synchronized with the electroencephalogram (EEG) signal and transmitted to the path recognition and planning module in real time.

[0095] Multi-source acquisition of robot gait data is achieved by installing six-axis MEMS inertial measurement units (IMUs) at the hip and knee joints of the lower limb rehabilitation robot, as well as at the center of the robot's chassis. Incremental encoders are also integrated into the robot's wheel hub motors and joint motors. The IMU sampling rate is set to ≥200Hz, and the encoder sampling rate is ≥1000Hz. The IMUs collect three-axis acceleration and three-axis angular velocity data during the robot's walking process, while the encoders collect the motor's rotation angle and speed data. Simultaneously, thin-film pressure sensors are installed on the robot's foot support structure to collect foot pressure distribution data during walking, which helps determine the gait phase (support phase, swing phase).

[0096] Gait feature extraction and feature vector generation involve fusing and processing acquired IMU, encoder, and pressure sensor data. A Kalman filter algorithm is used to calculate the posture of the IMU data, yielding real-time angles, angular velocities, and angular accelerations of the robot joints, as well as the robot's posture, position, and velocity. Based on the processed data, core gait feature parameters are extracted: spatiotemporal parameters (stride length, cadence, gait speed, walking cycle, and the proportion of support / swing phases); and kinematic parameters (real-time angles of the hip / knee joints, rate of change of joint angles, and range of motion). The extracted core features are normalized according to the walking cycle to generate standardized gait feature vectors, which are synchronized with EEG and EMG signals and transmitted in real-time to the path recognition and planning module.

[0097] In some embodiments, the perception of depth information via a depth camera includes: acquiring three-dimensional point cloud data of the site using the structured light or time-of-flight principle of the depth camera; removing outliers using voxel filtering; segmenting obstacle point clouds using Euclidean clustering algorithm; calculating the position coordinates, size, and distance of obstacles from the robot; and identifying the color or texture features of path boundaries using an edge detection algorithm to determine the direction and distance of path deviation.

[0098] This embodiment is the core implementation scheme for depth camera visual depth information perception in step S102, and it addresses the pain points of existing visual perception, such as susceptibility to environmental interference, low obstacle recognition accuracy, and inaccurate path deviation judgment. It acquires 3D point cloud data through structured light / time-of-flight principles, and achieves accurate obstacle detection and localization through filtering, noise reduction, and clustering segmentation. Edge detection algorithms are used to identify path boundaries and quantify path deviations, providing accurate environmental data for subsequent path planning, obstacle avoidance, and safety protection, thereby improving the robot's environmental adaptability and safety during autonomous movement.

[0099] Real-time acquisition of 3D point cloud data of the site is achieved using an RGB-D depth camera based on structured light or Time-of-Flight (ToF) principles. This camera is mounted on the front end of the lower limb rehabilitation robot, at a height of 0.8m-1.2m above the ground, with a pitch angle tilted downwards of 5°-10° to ensure coverage of the walking area 0.1m-8m in front of the robot and a horizontal field of view of 120°. The depth camera simultaneously acquires RGB color images and 3D depth point cloud data of the walking area at a frame rate of ≥30fps. The point cloud data includes the 3D spatial coordinates (X,Y,Z), RGB color information, and depth confidence of each pixel. For depth data with low confidence, a neighborhood interpolation algorithm is used to complete the data and eliminate depth holes.

[0100] Point cloud data preprocessing and denoising optimization involve preprocessing the acquired raw 3D point cloud data: First, through pass-through filtering, effective point clouds within a range of 0.1m-8m in front of the robot and 3m to the left and right are retained, while invalid point clouds outside the field of view are removed. Then, a voxel filtering algorithm is used, with the voxel grid size set to 5mm×5mm×5mm, to downsample the raw point cloud, reducing the amount of point cloud data and improving the processing speed while preserving the overall contour features of the point cloud. Finally, a statistical filtering algorithm is used to calculate the average distance from each point to its 50 neighboring points, removing outliers whose average distance exceeds the threshold, and eliminating invalid point clouds caused by environmental noise and shot noise, resulting in clean and effective point cloud data.

[0101] Obstacle point cloud segmentation and parameter calculation: For the preprocessed effective point cloud, Euclidean clustering algorithm is used for obstacle segmentation: taking each point in the point cloud as the core, a clustering distance threshold of 20mm is set, and adjacent points with a distance less than the threshold are divided into the same cluster to complete the point cloud segmentation of different obstacles; small clusters with fewer than 50 points are removed to eliminate interference from ground debris and noise; for each segmented obstacle point cloud cluster, its core parameters are calculated: position and size: using the minimum bounding box algorithm, the minimum bounding box of the obstacle point cloud is calculated, and the coordinates of the center of the box are taken as the three-dimensional position coordinates of the obstacle, while the outline dimensions (length, width, height) of the obstacle are obtained; relative distance: the Euclidean distance between the center of the obstacle and the center of the robot body, as well as the distance between the nearest point of the obstacle and the robot are calculated; motion trend: for the obstacle point cloud of continuous frames, the Kalman filter algorithm is used to track the position changes of the obstacle, calculate its motion speed and direction, and predict the motion trajectory; all obstacle parameter information is transmitted to the path recognition and planning module in real time.

[0102] Path boundary recognition and deviation judgment are performed using RGB color images acquired by a depth camera. An edge detection algorithm is used to identify path boundaries: First, the RGB images are preprocessed by grayscale conversion and Gaussian filtering. Then, the Canny edge detection algorithm is used to extract edge contours from the images. Combining preset path boundary features (such as the color and texture features of the site boundary lines, and the intersection of the wall and ground at the edge of the passage), the straight line contours of the path boundaries are extracted using Hough line transform to determine the path boundary lines on the left and right sides. Based on the site coordinate mapping table, the path boundary lines in the images are converted into actual boundary coordinates in a digital space coordinate system. Simultaneously, the real-time position and heading angle of the robot are obtained. The vertical distance between the robot center and the left and right path boundaries, as well as the angle between the robot's heading angle and the path centerline, are calculated to determine the direction (left / right deviation) and distance of the path deviation. The risk of exceeding the boundary is also assessed, and the path deviation data is transmitted to the path recognition and planning module in real time.

[0103] In some embodiments, the process of fusing the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and motion intention information to generate path adjustment instructions includes: establishing a multi-source data fusion framework; performing time-series alignment and weight allocation on the training route coordinates, obstacle positions, electromyographic intensity, gait characteristics, and motion intention levels using a Kalman filter algorithm; generating an obstacle avoidance path based on a preset obstacle avoidance algorithm; dynamically adjusting the target speed in conjunction with the motion intention intensity; and generating motor speed and steering angle instructions.

[0104] This embodiment is the core implementation scheme for multi-source data fusion and path adjustment instruction generation in step S103. It addresses the pain points of existing technologies, such as asynchronous multi-source data, chaotic decision priorities, and disconnect between path planning and user status. By building a multi-source data fusion framework, using the Kalman filter algorithm to achieve time-series alignment and adaptive weight allocation of multi-source data, and combining obstacle avoidance path generation algorithms with motion intent adaptation, it completes the standardized generation of path optimization and motor control instructions, constructing a complete decision-making chain of "data fusion - path planning - instruction generation." This ensures that the robot's autonomous walking not only meets rehabilitation training goals but also adapts to user intent and dynamic environmental changes.

[0105] The multi-source data fusion framework is constructed by building a distributed multi-source data fusion framework, which includes four layers: data access layer, time synchronization layer, weight allocation layer, and fusion decision layer. The data access layer respectively accesses the reference coordinate data of the training route, obstacle and path boundary data of visual perception, muscle strength characteristics of electromyography signals, robot gait feature vectors, and movement intention type and intensity level of EEG decoding, so as to achieve standardized access of multi-source data. All accessed data are timestamped using the unified hardware clock of the robot's main controller to ensure the consistency of the data's time reference.

[0106] The Kalman filter algorithm is used for time series alignment and optimal system state estimation. It aligns multi-source time series data and estimates the optimal state: First, a system state equation is established, with the robot's position coordinates, heading angle, walking speed, affected limb muscle strength, and movement intention intensity as system state variables; the training route reference coordinates, visually perceived obstacle positions, electromyographic intensity, gait characteristics, and movement intention level as observation variables. Through the prediction step of the Kalman filter, the system state at the current moment is predicted based on the system state at the previous moment. Through the update step, the predicted state is corrected by combining the multi-source observation data at the current moment, resulting in the optimal state estimate of the system at the current moment. Simultaneously, timestamp alignment is performed through Kalman filtering, unifying multi-source data with different sampling rates into a 100ms planning period, thus solving the time synchronization problem caused by different sensor sampling frequencies.

[0107] Adaptive weight allocation of multi-source data involves adaptive weight allocation of aligned multi-source data at the fusion decision layer: Preset base weights: 40% for motion intent information, 25% for training route baseline data, 20% for visual perception data, and 15% for a combination of electromyography (EMG) intensity and gait characteristics; Dynamic adjustment rules: When visual perception detects an obstacle distance less than the safety threshold (2m) or a risk of crossing the boundary, the weight of visual perception data increases to 35%, and the weight of the training route decreases to 10%, prioritizing safety; When EMG signals detect muscle fatigue or decreased muscle strength, the weight of EMG and gait characteristics increases to 25%, while the speed adjustment weight decreases, limiting the maximum walking speed; When EEG decodes an emergency braking or stop command, the weight of motion intent information increases to 100%, prioritizing the response to user-initiated commands; After weight allocation, the fused system decision parameters are obtained, including baseline trajectory, safety constraints, speed limit, steering restrictions, and posture correction parameters.

[0108] Obstacle avoidance path generation and trajectory smoothing optimization are based on the fused decision parameters, using either the artificial potential field method or the dynamic window method (DWA) to generate obstacle avoidance paths. Artificial potential field method: set up a gravitational potential field with the target point of the training route as the gravitational source and set up a repulsive potential field with obstacles and site boundaries as the repulsive source. By solving the resultant force direction of the gravitational and repulsive forces, a collision-free obstacle avoidance path for the robot is generated. Dynamic window method: In the robot's velocity space (linear velocity, angular velocity), a dynamic velocity window that meets the motor performance and safety constraints is generated. Trajectory simulation is performed for each set of velocities within the window. The target fit, obstacle avoidance safety, and smoothness of the trajectory are evaluated, and the trajectory with the highest score is selected as the optimal obstacle avoidance path. After generating the obstacle avoidance path, a cubic spline interpolation algorithm is used to smooth and optimize the path, ensuring the curvature of the trajectory is continuous and avoiding sudden turns and stops.

[0109] The motion parameter adaptation and control command generation, combined with the motion intent intensity level of EEG decoding, dynamically adjusts the robot's target motion parameters: based on the speed preference coefficient of the motion intent, the target walking speed is adjusted, with the speed adjustment range being 30%-120% of the baseline speed corresponding to the rehabilitation stage; based on the steering tendency coefficient and intent intensity level, the steering angle and turning radius are adjusted; based on electromyographic intensity and gait characteristics, stride length and stance phase duration are adjusted to adapt to the muscle strength of the affected limb; the optimized path trajectory and motion parameters are converted into control commands that can be recognized by the robot's actuators, including the target speed and steering angle of the left and right front wheel hub motors, the target rotation angle and stride parameters of the joint drive motors, and start / stop control commands; all control commands are standardized pulse width modulation (PWM) signals and CAN bus communication commands, which can be directly sent to the robot's execution terminal.

[0110] In some embodiments, controlling the lower limb rehabilitation robot to autonomously execute walking movements according to path adjustment instructions to complete rehabilitation training without human intervention includes: converting the speed and steering parameters in the path adjustment instructions into differential speed control signals for the front wheel hub motors; driving the motors of the lower limb rehabilitation robot to perform steering and speed adjustment movements through pulse width modulation technology; adaptively adjusting the support direction of the rear omnidirectional wheels of the lower limb rehabilitation robot; monitoring feedback data from depth cameras and electroencephalogram (EEG) signals; and triggering braking or path replanning procedures when the distance to an obstacle is detected to be less than a safety threshold or the movement intention command is changed.

[0111] This embodiment is the core implementation scheme for the robot's autonomous walking execution and closed-loop control in step S103. It addresses the pain points of existing rehabilitation robots, such as delayed execution response, low steering control accuracy, and lack of real-time feedback closed loops. Precise speed regulation and steering control are achieved through differential drive of the front wheel hub motors. A real-time feedback and PID closed-loop correction mechanism for execution status is established. Simultaneously, real-time monitoring and emergency response logic for visual and EEG signals are integrated to achieve fully closed-loop control and dynamic path replanning for autonomous walking. Ultimately, this completes rehabilitation training without human intervention, ensuring the accuracy, stability, and safety of walking control.

[0112] After receiving the path adjustment command, the robot's main controller first converts the target linear velocity and steering angle parameters in the command into differential speed control signals for the left and right hub motors of the front wheels. Based on the Ackerman steering principle, the target speed difference between the left and right hub motors is calculated according to the target steering angle, the robot's wheelbase, and the wheel track. During steering, the speed of the inner motor decreases and the speed of the outer motor increases, achieving precise steering through differential speed control. The target linear velocity is then converted into the target rated speed of the hub motors. The target speeds of the left and right motors are converted into pulse width modulation (PWM) drive signals with corresponding duty cycles. Simultaneously, speed and steering control commands are sent to the motor drivers via the CAN bus. The command update frequency is ≥100Hz to ensure real-time control.

[0113] After receiving PWM drive signals and control commands, the hub motor driver drives the left and right front wheel hub motors to execute corresponding speed and steering actions. Through differential coordination of the motors, the robot can achieve straight-line movement, acceleration, deceleration, left turn, and right turn. The robot's rear wheels use two omnidirectional wheels, which adaptively adjust the support direction according to the walking direction and steering action of the front wheels, ensuring the support stability and steering flexibility of the robot during walking. At the same time, the main controller converts the stride length and support phase duration parameters in the path adjustment command into control signals for the hip and knee joint drive motors, driving the joint motors to adjust gait parameters to adapt to the movement ability of the affected limb and complete lower limb rehabilitation assistive actions that match the walking movements.

[0114] During robot movement, the encoders of the hub motors and joint motors collect the actual speed and angle data of the motors in real time, while the IMU collects the robot's actual walking speed, heading angle, and position data in real time, and transmits them back to the main controller synchronously. The main controller compares the actual execution status with the target parameters of the path adjustment command, calculates the speed deviation, position deviation, and heading angle deviation, and adjusts the PWM drive signal of the motor in real time through the PID closed-loop control algorithm to correct the execution deviation, ensuring that the deviation between the robot's actual walking trajectory and the planned path is ≤±50mm, and the deviation between the actual speed and the target speed is ≤±5%, forming a full closed-loop control of "command issuance-execution-feedback-correction".

[0115] During the robot's autonomous walking process, the system monitors the visual perception data from the depth camera and the decoded feedback data from the electroencephalogram (EEG) signal in real time, with a period of 30ms: Visual data monitoring: Real-time monitoring of the closest distance between obstacles and the robot. When the distance to an obstacle is detected to be less than a preset safety threshold (1.5m in the early stage of rehabilitation, 1m in the middle stage, and 0.8m in the later stage), or when a collision between the obstacle and the robot's trajectory is predicted, a graded braking program is immediately triggered. The robot first decelerates to 30% of the baseline speed. If the distance continues to decrease, an emergency braking program is immediately triggered, and the joint motor movements are stopped. EEG signal monitoring: Real-time monitoring of the motor intention commands decoded from the EEG signal. When a change in the motor intention command is detected, the current execution command is immediately terminated, and the robot responds to the new intention command while simultaneously adjusting the path planning parameters. When an invalid or abnormal EEG signal is detected, a smooth braking program is immediately triggered to control the robot to decelerate to a stop.

[0116] When obstacle avoidance requirements are triggered, movement intention commands are changed, or path deviation exceeds a threshold, the system immediately triggers a path replanning program. Based on the latest multi-source fusion data, the optimal walking path and control commands are regenerated, and the robot continues to perform walking actions according to the new path. Throughout the entire rehabilitation training process, the robot autonomously completes path planning, walking control, obstacle avoidance, posture adjustment, and start-stop actions without manual operation or remote intervention. Alarm information is only sent when emergency braking or signal abnormality is triggered, ultimately completing lower limb rehabilitation training without human intervention.

[0117] In some embodiments, obtaining the training route corresponding to the user's current rehabilitation stage includes: automatically identifying the rehabilitation stage based on the rehabilitation stage identifier input by the user or through historical training data, and retrieving the corresponding route template from a preset route database: the rehabilitation stage includes the initial, middle, and late stages; the initial stage matches a first speed range and a straight path template, the middle stage matches a second speed range and a turning path template, and the late stage matches a third speed range and a path template containing slopes or complex curves, and dynamically fine-tuning the route parameters based on the user's real-time physiological signals; wherein, the first speed range is smaller than the second speed range, and the second speed range is smaller than the third speed range.

[0118] This embodiment is the core implementation scheme for training route acquisition and rehabilitation stage adaptation in step S101, and it fundamentally addresses the pain points of existing technology route templates being disconnected from the patient's rehabilitation cycle and lacking targeted rehabilitation training. By determining the user's rehabilitation stage through a dual-mode "manual input / automatic recognition," a graded route template library conforming to stroke rehabilitation medicine standards is established, achieving precise matching between training routes and the patient's rehabilitation capabilities. Simultaneously, it supports dynamic fine-tuning of route parameters based on real-time physiological signals, ensuring that the training routes not only comply with rehabilitation medicine standards but also adapt to the patient's real-time state, thereby improving the effectiveness and safety of rehabilitation training.

[0119] The system supports two rehabilitation stage determination modes to adapt to different usage scenarios: Manual input mode: Operators input the user's rehabilitation stage identifier or the user's Brunnstrom hemiplegia stage and Fugl-Meyer motor function score through the host computer interface, and the system automatically maps it to three rehabilitation stages: early, middle, and late. Automatic recognition mode: The system retrieves the user's historical training data, including historical training duration, walking speed, stride length, affected limb muscle strength, accuracy of motor intention recognition, training completion rate, and other core indicators. Through a pre-trained random forest classification model, it automatically identifies the user's current rehabilitation stage with an accuracy rate of ≥90%. The rehabilitation stage division conforms to the guidelines for stroke rehabilitation medicine: Early stage corresponds to Brunnstrom stages 1-3, where the patient has poor balance and can only complete simple straight-line walking; Middle stage corresponds to Brunnstrom stages 4-5, where the patient's spasticity is reduced and they have some balance and turning ability; Late stage corresponds to Brunnstrom stage 6, where the patient's motor function has basically recovered and they can adapt to complex walking environments.

[0120] The system has a pre-built structured database of preset routes, storing corresponding route templates according to rehabilitation stages. Each route template includes standardized parameters such as trajectory type, speed range, stride range, turning radius, training duration, and rehabilitation goals. Based on the determined rehabilitation stage, the system automatically retrieves the corresponding route template from the database: Early rehabilitation stage: Matching a straight path template within the first speed range (0.1km / h-0.3km / h), with a total path length of 10m-30m, no turning points, a stride length of 20cm-30cm, and a support phase duration of ≥60%. The rehabilitation goal is to establish a correct walking pattern and improve the weight-bearing capacity of the affected limb. Mid-rehabilitation stage: Matching a path template within the second speed range (0.3km / h-0.6km / h) that includes turns. The first course includes 90° / 180° fixed-radius turns with a turning radius ≥1.5m, a total course length of 30m-50m, and a stride length of 30cm-40cm. The rehabilitation goal is to improve steering control and walking stability. The second course uses a third speed range (0.6km / h-1.0km / h) with course templates containing slopes or complex curves. This includes S-shaped curves, continuous turns, and 3°-5° uphill / downhill sections with a turning radius ≥1m. The total course length is 50m-100m, and the stride length is 40cm-50cm. The rehabilitation goal is to improve walking adaptability in complex environments and restore daily walking function. The first speed range < the second speed range < the third speed range to ensure that the training difficulty is progressively matched to the patient's rehabilitation ability.

[0121] After retrieving the corresponding route template, the system dynamically fine-tunes the route parameters within a safe range based on the user's real-time physiological signals: it collects the user's EEG, EMG, and heart rate signals in real time, and extracts physiological characteristics such as the intensity of movement intention, the degree of muscle activation in the affected limb, muscle fatigue, and heart rate changes; when it detects that the user has high muscle fatigue, low intensity of movement intention, or a heart rate exceeding the safe range, it automatically lowers the baseline speed of the route, reduces stride length, shortens the total path length, and increases the turning radius to reduce training difficulty; when it detects that the user has high intensity of movement intention, good muscle activation, and a stable training process, it can appropriately raise the upper limit of speed and increase the path length to enhance training intensity; the fine-tuned route parameters do not exceed the safe threshold range of the corresponding rehabilitation stage.

[0122] After fine-tuning the parameters, the system binds the route template to the user's personal rehabilitation record, allowing users to customize and modify route parameters, including trajectory length, turning points, speed range, and slope, to adapt to different training venues such as home and rehabilitation institutions. At the same time, the route template and training parameters used in this session are stored in the user's historical training database for automatic identification and iterative optimization of routes in subsequent rehabilitation stages, achieving personalized training route adaptation throughout the entire rehabilitation cycle.

[0123] In some embodiments, this invention addresses the technical pain points of existing lower limb rehabilitation robots that rely on manual assistance and cannot autonomously adapt to diverse rehabilitation scenarios. Based on hybrid paradigm EEG signal acquisition, dual detection, and decoding technology, it adds a path recognition and planning module, a depth camera visual perception module, and a preset route storage and retrieval module. It also incorporates auxiliary correction functions for electromyography signals and gait signals of the affected limb. This constructs a closed-loop control logic of "multimodal physiological signal decoding (motor intention) + preset route matching + real-time depth camera perception + dynamic path adjustment + autonomous robot walking". This enables stroke hemiplegic patients to undergo lower limb rehabilitation training without human intervention in all scenarios, including at home, in rehabilitation institution corridors, and in dedicated rehabilitation areas, while ensuring walking safety and the effectiveness of rehabilitation training.

[0124] For example, the lower limb rehabilitation training system of the present invention consists of four parts: the original core module, the newly added functional module, the signal acquisition terminal, and the robot execution terminal. The structure and connection relationship of each part are as follows: The original core modules are: a hybrid paradigm EEG signal acquisition module, an EEG signal dual detection module, and an EEG signal decoding module. The EEG signal acquisition module is used to acquire the patient's hybrid EEG signals. The dual detection module is used to filter noise and extract features from the acquired EEG signals. The decoding module is used to quantify the extracted EEG features into the intensity of motor intention and to identify walking commands such as start, stop, and speed adjustment.

[0125] New functional modules include: a preset route storage and retrieval module, a depth camera visual perception module, and a path recognition and planning module. The preset route storage and retrieval module is used to pre-store training route data for different rehabilitation stages, such as straight lines, turns, and slope adaptation. The depth camera visual perception module is mounted on the robot body and is used to capture site markings, surrounding environmental obstacles, and scene boundaries. The path recognition and planning module is communicatively connected to the preset route storage and retrieval module, the depth camera visual perception module, the EEG signal decoding module, the EMG signal acquisition module, and the gait signal acquisition module, respectively, and is used to fuse multi-source data and output path adjustment commands.

[0126] Signal acquisition terminal: electromyography signal acquisition module and gait signal acquisition module; the electromyography signal acquisition module is attached to the muscle group of the patient's affected limb and is used to acquire electrical signals of muscle activity.

[0127] Robot execution terminal: The lower limb rehabilitation robot body adopts a motion configuration of front wheel hub motor + rear wheel universal wheel. It is connected to the path recognition and planning module to receive path adjustment commands and complete the autonomous adjustment of speed, stride and direction.

[0128] The core working principle of this invention is autonomous path planning and walking control driven by multi-source data fusion: the depth camera visual perception module captures site markers to complete spatial coordinate system calibration and establishes a mapping relationship between the physical site and the digital space; the EEG signal decoding module obtains the patient's active movement intention as the core control command for robot walking; the electromyography and gait signal acquisition module obtains the patient's limb state data to correct walking posture; the path recognition and planning module integrates preset route data, visual environment data, and multimodal physiological signal data to identify path deviations, obstacles, and scene boundaries in real time, dynamically generating a walking path adapted to the patient's rehabilitation needs, and ultimately driving the robot to achieve autonomous walking without human intervention.

[0129] The preparatory stage includes: ① Site marking: affixing highly recognizable flat reflective positioning marks at the four corners and geometric center of the rehabilitation site; ② Preset route import: importing matching training routes into the preset route storage and retrieval module according to the patient's rehabilitation stage.

[0130] During the coordinate system calibration phase, after the operator selects the target rehabilitation trajectory path, the system controls the depth camera's visual perception module to collect the feature information of the reflective markings on the site. Combined with the pre-input site size parameters, the coordinate system calibration and mapping from the physical site to the digital space is completed. The preset route is adjusted according to the actual scene to provide reference coordinates for subsequent path planning.

[0131] The multi-signal acquisition and decoding stage includes: ① EEG signal acquisition and decoding: The EEG signal acquisition module acquires the patient's mixed EEG signals in real time. After the noise is filtered out and features are extracted by the dual detection module, the decoding module quantifies the intensity of the movement intention and simultaneously identifies walking commands such as start, stop, acceleration, and deceleration; ② Auxiliary signal acquisition: The electromyography (EMG) signal acquisition module simultaneously acquires the EMG signals of the patient's affected limb, and the gait signal acquisition module acquires the gait data during the robot's walking process; ③ Visual environment acquisition: The depth camera visual perception module captures three-dimensional obstacles, path deviations, and scene boundary information in the field in real time and transmits them to the path recognition and planning module.

[0132] In the path fusion planning and dynamic adjustment stage, the path recognition and planning module receives preset route data, visual environment data, and multimodal physiological signal data, and performs the following operations: ① Path matching: compares the real-time path calibrated by vision with the preset route to identify the path deviation; ② Obstacle discrimination: locates and identifies the contours of visually captured three-dimensional obstacles and calculates the safe avoidance distance; ③ Intent fusion: combines the motor intent commands decoded by EEG to dynamically adjust the robot's walking speed, stride, and direction; ④ Posture correction: fine-tunes the robot's walking posture based on electromyography and gait signal data to ensure that the gait matches the muscle strength of the patient's affected limb.

[0133] During the autonomous walking and rehabilitation training phase, the robot receives control commands from the path recognition and planning module. It achieves steering and speed adjustment through differential control of the front wheel hub motor and adapts the steering action with the rear omnidirectional wheel to complete autonomous walking. During walking, if the depth camera detects new obstacles or path deviations, or the EEG decodes the patient's speed adjustment or stop commands, the system will adjust the path in real time without human intervention.

[0134] This method builds upon existing hybrid-paradigm EEG acquisition, dual-detection, and decoding technologies, adding a path recognition and planning module, a depth camera visual perception module, and a preset route storage and retrieval module. It constructs a closed-loop logic of "multimodal physiological signal decoding (motor intent) + preset route matching + real-time depth camera perception + dynamic path adjustment + autonomous robot walking." Specifically, preset routes for the rehabilitation scenario (such as straight lines, turns, and slope-adapted routes) are pre-imported. The depth camera captures the surrounding environment and path markers in real time, simultaneously decoding the patient's mixed EEG signals (quantifying the intensity of motor intent and recognizing walking commands such as start, stop, and speed adjustment). This assists in acquiring electromyographic signals and gait signals from the affected limb to correct walking posture. The path recognition module integrates preset route data with depth camera visual data to identify path deviations, obstacles, and rehabilitation scenario boundaries. The planning module combines the motor intent corresponding to the multimodal physiological signals to dynamically adjust walking speed, stride, and direction, ensuring that the robot's autonomous walking meets the patient's rehabilitation needs without additional human intervention, while simultaneously guaranteeing walking safety and the effectiveness of rehabilitation training.

[0135] The corresponding trajectory and location markers are as follows: Figures 3 to 13 As shown, in Figure 3 In this training program, an elliptical trajectory training diagram is provided to train patients' continuous turning coordination and dynamic balance abilities. This trajectory is a core training track for the transition from the mid-to-late stages of rehabilitation. L is the total length of the major axis of the elliptical trajectory, a key parameter determining the total distance walked in a single training session. A value of 5m-8m is recommended for the mid-rehabilitation period, and 8m-12m for the late-rehabilitation period; the larger the value, the longer the distance walked in a single training session. D is the total width of the minor axis of the elliptical trajectory, determining the turning span and difficulty of the trajectory. A smaller d value indicates a greater turning curvature and higher training difficulty. For example, a value of 2m-3m is recommended for the mid-rehabilitation period, and 1m-2m for the late-rehabilitation period. a is the length of one half of the major axis of the elliptical trajectory, a=L / 2, serving as the reference point for the major axis in the field coordinate system, corresponding to the positioning markers at both ends of the major axis. b is the width of one half of the minor axis of the elliptical trajectory, b=d / 2, serving as the reference point for the minor axis in the field coordinate system, corresponding to the positioning markers at both ends of the minor axis. The trajectory includes a continuously and gradually changing turning curvature without sharp turning points. It can train patients' ability to transfer their center of gravity and coordinate their affected limbs during continuous turning under low-impact conditions, avoiding the risk of balance instability caused by sharp right-angle turns.

[0136] exist Figure 4 The training program provides a circular trajectory training diagram for enhancing patients' omnidirectional turning control, constant curvature turning ability, and closed-loop dynamic balance ability. It serves as an advanced training trajectory in the later stages of rehabilitation. L is the total diameter of the circular trajectory, determining its overall size and turning radius. A smaller L value results in a smaller turning radius and higher turning training difficulty. A value of 3m to 6m is recommended for the later stages of rehabilitation. d is the total diameter of the circular trajectory; here, d=L to ensure the trajectory is a perfect circle. This serves as the positioning basis for the circumferential reference point when calibrating the field coordinate system. Four to eight path verification markers can be evenly distributed on the circumference. Through a continuous closed-loop trajectory with constant curvature, the program trains the patient's weight-bearing capacity of the affected limb and continuous center of gravity control during 360° omnidirectional turning. It is a core reinforcement training trajectory for patients before returning to complex daily walking scenarios.

[0137] exist Figure 5 In this training program, a square trajectory training diagram is provided for comprehensive training of patients' fixed-angle right-angle turns, start-stop control, and straight-line stability. This is the core training trajectory in the mid-stage of rehabilitation. L is the total length of the long side of the square trajectory, the core dimension of the straight-line segment, determining the distance of a single straight-line training session. A value of 4m~6m is recommended for the mid-stage of rehabilitation, corresponding to the positioning markers along the long axis of the field. W is the total width of the short side of the square trajectory, the dimension of the short-side straight-line segment. Typically, W=L forms a square trajectory, but it can also be adjusted to a rectangular trajectory depending on the field. A value of 4m~6m is recommended for the mid-stage of rehabilitation, corresponding to the positioning markers along the short axis of the field. d is the radius of the rounded transition section of the square trajectory, the arc transition dimension at the four right angles, used to avoid patient instability due to sharp right-angle turns. A value of 0.5m~1m is recommended for the early to mid-stages of rehabilitation, and can be reduced to 0.2m~0.5m in the later stages. The smaller the d value, the closer the turn is to a right angle, and the higher the training difficulty. 'a' is the length of the single-sided semi-axis of the longer side of the square trajectory, a = L / 2, which is the basis for positioning the long side reference point in the site coordinate system. 'b' is the width of the single-sided semi-axis of the shorter side of the square trajectory, b = W / 2, which is the basis for positioning the shorter side reference point in the site coordinate system.

[0138] The trajectory includes a straight section and a 90° fixed-angle turning section, which can simultaneously train the patient's straight walking stability and right-angle turning control ability, perfectly matching the rehabilitation progression goal of patients in the middle stage of rehabilitation: "from straight walking to turning walking".

[0139] exist Figure 6 The system provides a schematic diagram of a straight reciprocating trajectory for training, which is used for basic training in establishing a patient's basic straight walking pattern, weight-bearing capacity of the affected limb, and start-stop control. It is a core basic training trajectory in the early stage of rehabilitation.

[0140] L is the total length of a single straight segment of the reciprocating linear trajectory, and it is the core dimension for a single straight-line training session. A value of 5m to 10m is recommended in the early stages of rehabilitation, and can be extended to 10m to 20m in the middle stages. The larger the value, the longer the distance of a single straight-line training session.

[0141] W represents the lateral spacing of the reciprocating straight path, which is the lateral offset distance of the round-trip walking route. It ensures that the round-trip routes do not overlap spatially and is compatible with the width of the rehabilitation area's passageway. A value of 0.8m to 1.2m is recommended.

[0142] D is the effective width of the straight reciprocating trajectory, which is the effective lateral passage width of a single straight walking route. It matches the width of the lower limb rehabilitation robot body and reserves adjustment space for path deviation. It is the core benchmark parameter for robot path deviation recognition. A value of 0.6m to 0.8m is recommended.

[0143] The trajectory has no turning movements, only includes straight walking, starting and stopping, and reversing direction. It has the lowest training difficulty and can help hemiplegic patients establish a correct straight walking pattern, reduce the risk of balance instability, and is the first choice trajectory for patients to start lower limb rehabilitation walking training.

[0144] exist Figure 7 The system provides a figure-eight trajectory training diagram for advanced reinforcement training of patients' bidirectional continuous turning, dynamic shift of center of gravity, and adaptation to complex paths. It is the highest-level training trajectory in the later stage of rehabilitation.

[0145] L is the total length of the major axis of a single loop in the figure-eight trajectory. It is the major axis dimension of a single closed loop and determines the walking distance of a single loop and the total training length. A value of 4m to 6m is recommended in the later stages of rehabilitation.

[0146] W is the total lateral width of the double loops in the figure-eight trajectory, representing the total lateral span of the two closed loops. Typically, W is set to L to ensure a symmetrical structure between the two closed loops. A value of 4m to 6m is recommended for later stages of rehabilitation.

[0147] D is the width of the transition segment at the intersection of the figure-eight trajectory, which is the effective passage width at the intersection of two closed loops. It determines the accuracy requirement for turning at the intersection. The smaller the value of d, the higher the turning accuracy requirement and the greater the training difficulty. A value of 0.8m to 1.2m is recommended.

[0148] The trajectory includes bidirectional, continuously varying curvature turning movements, requiring seamless switching of turning direction at intersections. It comprehensively trains patients' bidirectional turning control, dynamic balance, and ability to adapt to complex paths, serving as the final training trajectory for patients before returning to their homes and communities for daily walking.

[0149] Figures 8 to 12 Series and Figures 3 to 7The series of trajectories correspond one-to-one, and all the markers are the provided positioning markers. The core function is to complete the coordinate system calibration of the physical site and digital space, providing a spatial reference for the robot's real-time path matching, deviation recognition, and autonomous walking.

[0150] Figure 8 Is with Figure 3 The site path marking layout diagram that accompanies the elliptical trajectory is used for coordinate system calibration, continuous turning path matching, and real-time verification of walking deviation during elliptical trajectory training.

[0151] At the two endpoints of the major axis of the ellipse (corresponding to Figure 3 (the endpoints of the a-axis), the two endpoints of the minor axis (corresponding to) Figure 3 Five reference positioning markers are arranged, including the endpoint of the b-axis and the geometric center of the ellipse, forming the "four corners + center" positioning reference matrix in the technical disclosure document, thus completing the coordinate system mapping between the physical site and the digital space.

[0152] Along the circumference of the elliptical trajectory, evenly distribute 8 to 12 path guide signs, with a spacing of 2 to 3 meters between the signs, and ensure that the arrangement outlines perfectly match. Figure 3 The L (major axis) and d (minor axis) dimensions are used for real-time identification of path deviations during robot movement, with a deviation verification accuracy of ±10mm. The total length of the major axis and the total width of the minor axis of the marking arrangement are... Figure 3 The L and d values ​​are completely consistent, ensuring that the marked trajectory coincides with the preset training trajectory in a 1:1 ratio.

[0153] Figure 9 Is with Figure 4 The site path marking layout diagram for circular trajectory is used for coordinate system calibration, closed-loop path matching, and continuous steering accuracy verification during circular trajectory training.

[0154] A central positioning marker is placed at the geometric center of the circular trajectory. Reference positioning markers are then placed at 0°, 90°, 180°, and 270° along the circumference of the circular trajectory, forming a "center + four-directional" positioning reference matrix. The arrangement radius is... Figure 4 The L / 2 (circle radius) is perfectly matched, and the coordinate system calibration is completed.

[0155] Between the four reference positioning markers, four to eight path guidance markers are evenly added along the circumference, for a total of eight to twelve markers on the entire circumference. The marker spacing is 1m to 2m. This is used for real-time path deviation identification during the robot's 360° continuous turning process to ensure the trajectory fit of the closed-loop walking.

[0156] The diameter of all circular markings is... Figure 4 The L values ​​are completely consistent, ensuring that the marked trajectory coincides with the preset circular training trajectory.

[0157] Figure 10 Is with Figure 5 The site path marking layout diagram accompanying the square trajectory is used for coordinate system calibration, straight segment path matching, and right-angle turning accuracy verification during square trajectory training.

[0158] One reference positioning marker is placed at each of the four vertices of the square trajectory, and one center positioning marker is placed at the geometric center of the square, forming a "four corners + center" 5-point positioning reference matrix as described in the technical disclosure document. The spacing between the four vertices markers is... Figure 5 The L (long side) and W (short side) are perfectly matched, thus completing the coordinate system calibration.

[0159] At the midpoint of each of the four sides of the square trajectory, one straight-line verification marker is placed for identifying straight-line path deviations; at the four rounded transition sections (corresponding to...) Figure 5 Each corner (d-radius) is marked with 1-2 steering verification markers for path matching and accuracy verification during right-angle turns. The dimensions of the long and short sides of the markers are... Figure 5 The L and W are completely consistent, and the arrangement radius of the rounded transition section markings is the same as... Figure 5 The d is completely identical.

[0160] Figure 11 Is with Figure 6 The site path marking layout diagram that accompanies the straight reciprocating trajectory is used for coordinate system calibration, straight path deviation identification, and reciprocating reversal point positioning during straight trajectory training. It is the most commonly used marking layout scheme in the early stage of rehabilitation.

[0161] One reference positioning marker is placed at the starting point, ending point, and reversal point of the straight reciprocating trajectory, and one boundary positioning marker is placed at each end of the left and right boundaries of the trajectory, forming a 5-point positioning reference matrix. The arrangement range is [missing information]. Figure 6 The L (length of a single straight line segment) and W (reciprocating horizontal spacing) are perfectly matched, thus completing the coordinate system calibration.

[0162] Path verification markers: Along the length of a single straight track, one path guidance marker is placed every 2 meters. The markers on the left and right sides form the valid path boundary for straight walking, with the left and right spacing being the same as... Figure 6 The d (effective path width) is completely consistent, used for real-time identification of lateral deviations in the robot's path, ensuring straight-line walking accuracy. The straight length, reciprocating spacing, and path width of the marker placement are consistent with... Figure 6 The L, W, and d correspond one-to-one.

[0163] Figure 12 Is with Figure 7 The accompanying site path marking layout diagram for the figure-eight trajectory is used for coordinate system calibration, complex turning path matching, and intersection turning accuracy verification during figure-eight trajectory training.

[0164] One center positioning marker is placed at the geometric center of each of the two closed loops of the figure-eight trajectory and at the geometric center of the overall trajectory (intersection point). One reference positioning marker is placed at each of the two endpoints of the major axis and the two endpoints of the minor axis of the figure-eight trajectory, forming a 7-point positioning reference matrix. The arrangement range is [not specified in the original text]. Figure 7 The L (major axis of the single ring) and W (total width of the double ring) are perfectly matched, thus completing the coordinate system calibration.

[0165] Along the two closed loops of the figure-eight trajectory, 6-8 path guide signs are evenly distributed on each side; at the intersection (corresponding to...) Figure 7 Two steering verification markers are placed along the intersection width (d), with a total of 14-18 markers along the entire trajectory. These are used for real-time path deviation identification during continuous bidirectional steering of the robot, ensuring the fit of complex trajectories. The major axis of the single ring, the total width of the double rings, and the intersection width of the markers are determined by... Figure 7 The L, W, and d correspond one-to-one.

[0166] Please see Figure 13 As shown, Figure 13 This is a schematic diagram of the structure of a self-propelled path recognition and planning system 200 for a lower limb rehabilitation robot provided in this application embodiment. The self-propelled path recognition and planning system 200 for a lower limb rehabilitation robot is used to execute the steps of the self-propelled path recognition and planning method for a lower limb rehabilitation robot shown in the above embodiments. The self-propelled path recognition and planning system 200 for a lower limb rehabilitation robot can be a single server or a server cluster, or it can be a terminal, such as a handheld terminal, a laptop computer, a wearable device, or a robot.

[0167] like Figure 13 As shown, the self-walking path recognition and planning system 200 for lower limb rehabilitation robots includes: The route acquisition unit 201 is used to acquire the training route corresponding to the current rehabilitation stage of the user, set multiple positioning markers in the rehabilitation site, and establish a coordinate system mapping between the physical site and the digital space based on the positioning markers and the site parameters corresponding to the rehabilitation site, so as to adjust the training route. The signal acquisition unit 202 is used to acquire the mixed EEG signal corresponding to the user, decode and obtain the movement intention information corresponding to the mixed EEG signal; acquire the electromyographic signal of the affected limb corresponding to the user and the robot gait data corresponding to the lower limb rehabilitation robot; and perceive depth information through a depth camera, wherein the depth information includes at least site obstacle information, path deviation information and boundary information. The training completion unit 203 is used to integrate the training route, visual perception data, electromyographic signals of the affected limb, robot gait data and movement intention information to generate path adjustment instructions; and control the lower limb rehabilitation robot to autonomously execute walking movements according to the path adjustment instructions to complete rehabilitation training without human intervention.

[0168] In some embodiments, establishing a coordinate system mapping between the physical site and the digital space based on the location markers and the site parameters corresponding to the rehabilitation site includes: acquiring image feature information of the location markers through a depth camera, identifying the corner points or contour features of the location markers using an image processing algorithm, and, in conjunction with the pre-input site length, width, and height parameters, establishing a one-to-one mapping relationship between each location point of the physical site and the digital space coordinate system based on the principle of perspective transformation and coordinate transformation algorithm, and generating a site coordinate mapping table.

[0169] In some embodiments, adjusting the training route includes: matching the coordinates of key nodes of the preset training route with the actual site coordinates in the coordinate system mapping table, calculating the positional and angular deviations of the route nodes, interpolating and correcting the deviation nodes using a path smoothing algorithm, or adjusting the curvature of the route based on a Bezier curve fitting algorithm, so that the training route conforms to the spatial layout of the actual site.

[0170] In some embodiments, the decoding to obtain the motion intention information corresponding to the mixed EEG signal includes: sequentially performing bandpass filtering on the acquired mixed EEG signal to remove noise, extracting the energy spectrum features of the characteristic frequency band through wavelet transform or principal component analysis, and using support vector machine or convolutional neural network classification algorithms to identify the type of motion intention and quantify the level of intention intensity.

[0171] In some embodiments, the acquisition of electromyographic signals of the affected limb corresponding to the user and robot gait data corresponding to the lower limb rehabilitation robot includes: acquiring electromyographic signals through surface electrodes attached to the muscle groups of the affected limb, and extracting the root mean square value or average absolute value as muscle activity intensity features after amplification, rectification and normalization; acquiring gait data through inertial sensors installed on the robot joints, extracting stride length, stride frequency and joint angle change rate, and generating gait feature vectors.

[0172] In some embodiments, the perception of depth information via a depth camera includes: acquiring three-dimensional point cloud data of the site using the structured light or time-of-flight principle of the depth camera; removing outliers using voxel filtering; segmenting obstacle point clouds using Euclidean clustering algorithm; calculating the position coordinates, size, and distance of obstacles from the robot; and identifying the color or texture features of path boundaries using an edge detection algorithm to determine the direction and distance of path deviation.

[0173] In some embodiments, the process of fusing the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and motion intention information to generate path adjustment instructions includes: establishing a multi-source data fusion framework; performing time-series alignment and weight allocation on the training route coordinates, obstacle positions, electromyographic intensity, gait characteristics, and motion intention levels using a Kalman filter algorithm; generating an obstacle avoidance path based on a preset obstacle avoidance algorithm; dynamically adjusting the target speed in conjunction with the motion intention intensity; and generating motor speed and steering angle instructions.

[0174] In some embodiments, controlling the lower limb rehabilitation robot to autonomously execute walking movements according to path adjustment instructions to complete rehabilitation training without human intervention includes: converting the speed and steering parameters in the path adjustment instructions into differential speed control signals for the front wheel hub motors; driving the motors of the lower limb rehabilitation robot to perform steering and speed adjustment movements through pulse width modulation technology; adaptively adjusting the support direction of the rear omnidirectional wheels of the lower limb rehabilitation robot; monitoring feedback data from depth cameras and electroencephalogram (EEG) signals; and triggering braking or path replanning procedures when the distance to an obstacle is detected to be less than a safety threshold or the movement intention command is changed.

[0175] In some embodiments, obtaining the training route corresponding to the user's current rehabilitation stage includes: automatically identifying the rehabilitation stage based on the rehabilitation stage identifier input by the user or through historical training data, and retrieving the corresponding route template from a preset route database: the rehabilitation stage includes the initial, middle, and late stages; the initial stage matches a first speed range and a straight path template, the middle stage matches a second speed range and a turning path template, and the late stage matches a third speed range and a path template containing slopes or complex curves, and dynamically fine-tuning the route parameters based on the user's real-time physiological signals; wherein, the first speed range is smaller than the second speed range, and the second speed range is smaller than the third speed range.

[0176] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the self-walking path recognition and planning system and its modules for lower limb rehabilitation robots described above can be found in the corresponding contents of the various embodiments of the self-walking path recognition and planning method for lower limb rehabilitation robots, and will not be repeated here.

[0177] The aforementioned method for self-walking path recognition and planning for lower limb rehabilitation robots can be implemented as a computer program, which can be used in various ways, such as... Figure 13 It runs on the device shown.

[0178] Please see Figure 14 , Figure 14This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0179] The storage medium may store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any self-walking path recognition and planning method for lower limb rehabilitation robots.

[0180] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0181] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any self-walking path recognition and planning method for lower limb rehabilitation robots.

[0182] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 14 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0183] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0184] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: Obtain the training route corresponding to the user's current rehabilitation stage, set multiple positioning markers in the rehabilitation site, and establish a coordinate system mapping between the physical site and the digital space based on the positioning markers and the site parameters corresponding to the rehabilitation site, so as to adjust the training route. Collect mixed EEG signals corresponding to the user, decode and obtain the movement intention information corresponding to the mixed EEG signals; collect electromyography signals of the affected limb corresponding to the user and robot gait data corresponding to the lower limb rehabilitation robot; perceive depth information through a depth camera, the depth information including at least site obstacle information, path deviation information and boundary information; By integrating the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and movement intention information, a path adjustment command is generated; the lower limb rehabilitation robot is controlled to autonomously execute walking movements according to the path adjustment command, thus completing rehabilitation training without human intervention.

[0185] In some embodiments, establishing a coordinate system mapping between the physical site and the digital space based on the location markers and the site parameters corresponding to the rehabilitation site includes: acquiring image feature information of the location markers through a depth camera, identifying the corner points or contour features of the location markers using an image processing algorithm, and, in conjunction with the pre-input site length, width, and height parameters, establishing a one-to-one mapping relationship between each location point of the physical site and the digital space coordinate system based on the principle of perspective transformation and coordinate transformation algorithm, and generating a site coordinate mapping table.

[0186] In some embodiments, adjusting the training route includes: matching the coordinates of key nodes of the preset training route with the actual site coordinates in the coordinate system mapping table, calculating the positional and angular deviations of the route nodes, interpolating and correcting the deviation nodes using a path smoothing algorithm, or adjusting the curvature of the route based on a Bezier curve fitting algorithm, so that the training route conforms to the spatial layout of the actual site.

[0187] In some embodiments, the decoding to obtain the motion intention information corresponding to the mixed EEG signal includes: sequentially performing bandpass filtering on the acquired mixed EEG signal to remove noise, extracting the energy spectrum features of the characteristic frequency band through wavelet transform or principal component analysis, and using support vector machine or convolutional neural network classification algorithms to identify the type of motion intention and quantify the level of intention intensity.

[0188] In some embodiments, the acquisition of electromyographic signals of the affected limb corresponding to the user and robot gait data corresponding to the lower limb rehabilitation robot includes: acquiring electromyographic signals through surface electrodes attached to the muscle groups of the affected limb, and extracting the root mean square value or average absolute value as muscle activity intensity features after amplification, rectification and normalization; acquiring gait data through inertial sensors installed on the robot joints, extracting stride length, stride frequency and joint angle change rate, and generating gait feature vectors.

[0189] In some embodiments, the perception of depth information via a depth camera includes: acquiring three-dimensional point cloud data of the site using the structured light or time-of-flight principle of the depth camera; removing outliers using voxel filtering; segmenting obstacle point clouds using Euclidean clustering algorithm; calculating the position coordinates, size, and distance of obstacles from the robot; and identifying the color or texture features of path boundaries using an edge detection algorithm to determine the direction and distance of path deviation.

[0190] In some embodiments, the process of fusing the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and motion intention information to generate path adjustment instructions includes: establishing a multi-source data fusion framework; performing time-series alignment and weight allocation on the training route coordinates, obstacle positions, electromyographic intensity, gait characteristics, and motion intention levels using a Kalman filter algorithm; generating an obstacle avoidance path based on a preset obstacle avoidance algorithm; dynamically adjusting the target speed in conjunction with the motion intention intensity; and generating motor speed and steering angle instructions.

[0191] In some embodiments, controlling the lower limb rehabilitation robot to autonomously execute walking movements according to path adjustment instructions to complete rehabilitation training without human intervention includes: converting the speed and steering parameters in the path adjustment instructions into differential speed control signals for the front wheel hub motors; driving the motors of the lower limb rehabilitation robot to perform steering and speed adjustment movements through pulse width modulation technology; adaptively adjusting the support direction of the rear omnidirectional wheels of the lower limb rehabilitation robot; monitoring feedback data from depth cameras and electroencephalogram (EEG) signals; and triggering braking or path replanning procedures when the distance to an obstacle is detected to be less than a safety threshold or the movement intention command is changed.

[0192] In some embodiments, obtaining the training route corresponding to the user's current rehabilitation stage includes: automatically identifying the rehabilitation stage based on the rehabilitation stage identifier input by the user or through historical training data, and retrieving the corresponding route template from a preset route database: the rehabilitation stage includes the initial, middle, and late stages; the initial stage matches a first speed range and a straight path template, the middle stage matches a second speed range and a turning path template, and the late stage matches a third speed range and a path template containing slopes or complex curves, and dynamically fine-tuning the route parameters based on the user's real-time physiological signals; wherein, the first speed range is smaller than the second speed range, and the second speed range is smaller than the third speed range.

[0193] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the steps of the self-walking path recognition and planning method for a lower limb rehabilitation robot as provided in any embodiment of this application.

[0194] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0195] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for self-walking path recognition and planning for a lower limb rehabilitation robot, characterized in that, include: Obtain the training route corresponding to the user's current rehabilitation stage, set multiple positioning markers in the rehabilitation site, and establish a coordinate system mapping between the physical site and the digital space based on the positioning markers and the site parameters corresponding to the rehabilitation site, so as to adjust the training route. Collect mixed EEG signals corresponding to the user, decode and obtain the movement intention information corresponding to the mixed EEG signals; collect electromyography signals of the affected limb corresponding to the user and robot gait data corresponding to the lower limb rehabilitation robot; perceive depth information through a depth camera, the depth information including at least site obstacle information, path deviation information and boundary information; By integrating the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and movement intention information, a path adjustment command is generated; the lower limb rehabilitation robot is controlled to autonomously execute walking movements according to the path adjustment command, thus completing rehabilitation training without human intervention.

2. The method according to claim 1, characterized in that, The step of establishing a coordinate system mapping between the physical site and the digital space based on the location identifier and the site parameters corresponding to the rehabilitation site includes: By acquiring image feature information of positioning markers using a depth camera, and using image processing algorithms to identify the corner points or contour features of the positioning markers, combined with the pre-input site length, width and height parameters, a one-to-one mapping relationship between each positioning point of the physical site and the digital space coordinate system is established based on the principle of perspective transformation and coordinate transformation algorithm, and a site coordinate mapping table is generated.

3. The method according to claim 2, characterized in that, The adjustment of the training route includes: The key node coordinates of the preset training route are matched with the actual site coordinates in the coordinate system mapping table. The positional and angular deviations of the route nodes are calculated. The deviation nodes are corrected by interpolation through a path smoothing algorithm, or the curvature of the route is adjusted based on a Bézier curve fitting algorithm, so that the training route conforms to the spatial layout of the actual site.

4. The method according to claim 1, characterized in that, The decoding process for obtaining the motor intention information corresponding to the mixed EEG signals includes: The acquired mixed EEG signals are sequentially bandpass filtered to remove noise. Energy spectrum features of characteristic frequency bands are extracted through wavelet transform or principal component analysis. Support vector machine or convolutional neural network classification algorithms are used to identify the type of motion intent and quantify the level of intent intensity.

5. The method according to claim 1, characterized in that, The collection of electromyographic signals of the affected limb corresponding to the user and gait data of the lower limb rehabilitation robot includes: Electromyographic signals are collected by surface electrodes attached to the muscle groups of the affected limb. After amplification, rectification and normalization, the root mean square value or mean absolute value is extracted as the muscle activity intensity feature. Gait data is collected by inertial sensors installed on the robot's joints, and stride length, stride frequency, and joint angle change rate are extracted to generate gait feature vectors.

6. The method according to claim 1, characterized in that, The method of sensing depth information through a depth camera includes: The site's 3D point cloud data is acquired using structured light or time-of-flight principles from a depth camera. Outliers are removed by voxel filtering, and Euclidean clustering is used to segment the obstacle point cloud. The position coordinates, size, and distance of the obstacles from the robot are then calculated. Edge detection algorithms are used to identify the color or texture features of path boundaries to determine the direction and distance of path deviation.

7. The method according to claim 1, characterized in that, The process of integrating the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and movement intention information to generate path adjustment instructions includes: A multi-source data fusion framework was established, and the Kalman filter algorithm was used to perform time-series alignment and weight allocation of training route coordinates, obstacle positions, electromyographic intensity, gait features and movement intention levels. The obstacle avoidance path is generated based on a preset obstacle avoidance algorithm, the target speed is dynamically adjusted in combination with the intensity of motion intent, and motor speed and steering angle commands are generated.

8. The method according to claim 1, characterized in that, The control of the lower limb rehabilitation robot to autonomously execute walking movements according to path adjustment instructions, and to complete rehabilitation training without human intervention, includes: The speed and steering parameters in the path adjustment command are converted into differential speed control signals for the front wheel hub motor. The motor of the lower limb rehabilitation robot is driven to perform steering and speed adjustment actions through pulse width modulation technology. The rear omnidirectional wheel of the lower limb rehabilitation robot adaptively adjusts the support direction. It monitors feedback data from depth cameras and EEG signals, and triggers braking or path replanning procedures when an obstacle is detected to be less than a safety threshold or when the movement intention command is changed.

9. The method according to claim 1, characterized in that, The step of obtaining the training route corresponding to the user's current rehabilitation stage includes: The rehabilitation stage is automatically identified based on the user's input of the rehabilitation stage identifier or through historical training data, and the corresponding route template is retrieved from the preset route database: the rehabilitation stage includes the initial, middle and late stages; the initial stage matches the first speed range and straight path template, the middle stage matches the second speed range and turning path template, and the late stage matches the third speed range and path template containing slope or complex curves, and the route parameters are dynamically fine-tuned according to the user's corresponding real-time physiological signals. Wherein, the first speed range is smaller than the second speed range, and the second speed range is smaller than the third speed range.

10. A self-propelled path recognition and planning system for a lower limb rehabilitation robot, characterized in that, The method applied to any one of claims 1-9 includes: The route acquisition unit is used to acquire the training route corresponding to the current rehabilitation stage of the user, set multiple positioning markers in the rehabilitation site, and establish a coordinate system mapping between the physical site and the digital space based on the positioning markers and the site parameters corresponding to the rehabilitation site, so as to adjust the training route. The signal acquisition unit is used to acquire the user's corresponding mixed EEG signal, decode and obtain the movement intention information corresponding to the mixed EEG signal; acquire the user's corresponding affected limb electromyography signal and the robot gait data corresponding to the lower limb rehabilitation robot; and perceive depth information through a depth camera, wherein the depth information includes at least site obstacle information, path deviation information and boundary information. The training completion unit is used to integrate the training route, visual perception data, electromyographic signals of the affected limb, robot gait data, and movement intention information to generate path adjustment instructions; and control the lower limb rehabilitation robot to autonomously execute walking movements according to the path adjustment instructions to complete rehabilitation training without human intervention.

Images