A gesture recognition and gesture evaluation method based on a wearable device

Abstract

The application relates to a posture recognition and posture evaluation method based on a wearable device and relates to the field of human posture detection. The method solves the problems of strong subjectivity, insufficient data dimension and limited evaluation accuracy of traditional posture evaluation which depends on manual observation or a single sensor. The method is as follows: a plurality of parts of an inertial measurement unit and a flexible plantar pressure sensor array are deployed and synchronously collected, time synchronization and other treatments are carried out, time windows of each action are determined by taking a sudden change point of a plantar pressure sequence as an event anchor point, a fusion feature vector is constructed, a deep evaluation is carried out in combination with a fusion feature vector for identifying a motion posture and a preset biomechanical model, differences between an action of a wearer and a standard action or a personal action feature baseline are quantified, and finally a personalized improvement report is generated. The method can be applied to the fields of football, volleyball, tennis and other sports evaluation fields, the fields of posture evaluation, motion control and intelligent manufacturing of industrial robots and man-machine cooperation.

Description

Technical Field

[0001] This invention relates to the fields of wearable device technology and human posture detection, and particularly to the field of posture detection technology for basketball. It can be applied to the fields of posture recognition and motion control for humanoid robots. Background Technology

[0002] Basketball is a highly dynamic, non-periodic sport with complex movements, and an athlete's posture control directly impacts performance and injury risk. Therefore, during basketball player training, it's essential to identify various movement postures to assess training status and results.

[0003] In existing technologies, the recognition and evaluation of movement postures mainly rely on manual observation or data collection from a single sensor, and manual evaluation, which has the drawback of strong subjectivity. The accuracy of the evaluation depends on the coach's experience. Using data collected from a single sensor for posture recognition and evaluation has problems such as limited evaluation accuracy due to insufficient data measurement dimensions. For example, analyzing measurement data obtained solely from an inertial measurement unit (IMU) device can measure the athlete's shooting angle, but cannot perceive the force exerted during the shot; data measured solely from a plantar pressure sensor can analyze the athlete's footwork during movement, but cannot obtain information about the athlete's upper body posture. Therefore, existing technologies cannot objectively, comprehensively, and accurately reflect the athlete's full-body coordinated movement posture during movement. Posture data obtained based on existing technologies cannot provide an objective evaluation and analysis of basketball players' movement postures, and thus cannot support training guidance for athletes. Summary of the Invention

[0004] This invention solves the problems of traditional attitude assessment relying on manual observation or a single sensor, which has strong subjectivity, insufficient data dimensions, and limited assessment accuracy.

[0005] The technical solution provided by this invention: A posture recognition method based on a wearable device, the wearable device comprising multiple inertial measurement units and two flexible foot pressure sensor arrays; the multiple inertial measurement units include a head inertial measurement unit for detecting the wearer's head posture, a torso inertial measurement unit for detecting the wearer's torso posture, and eight limb inertial measurement units respectively for the movement postures of the two upper arms, the two forearms, the two thighs, and the two lower legs. The posture recognition method is as follows: during the wearer's movement: The attitude data of each inertial measurement unit and the multi-channel plantar pressure data output by each flexible pressure sensor array are collected in real time and synchronously. Using the sequence mutation points of multi-channel plantar pressure data as event anchors, we can detect take-off, landing, sudden stop or turning action events, and further determine the time window of each action event; For each posture data collected within the time window, statistical and temporal features including linear acceleration, angular velocity, posture angle, and joint angle are extracted. At the same time, features of multi-channel plantar pressure data collected synchronously with each posture data are extracted. All statistical features and time-domain features within the aforementioned time window are fused with the features of the multi-channel plantar pressure data to construct a fused feature vector, with each vector corresponding to an action primitive. The fused feature vectors are input into a pre-trained machine learning model to identify basketball action types, including shooting, landing cushioning, and crossover dribbling.

[0006] Furthermore, in one embodiment of the present invention, after real-time synchronous acquisition of attitude data from each inertial measurement unit and multi-channel plantar pressure data output from each flexible pressure sensor array, the method further includes multi-source sensor data fusion of the acquired data. This fusion refers to spatiotemporal alignment of all attitude data and multi-channel plantar pressure data acquired at the same time, including: Time alignment: Synchronize all posture data with multi-channel plantar pressure data in time. Spatial alignment of attitude data, noise suppression and coordinate system integration of all attitude data to complete attitude data fusion; All multi-channel plantar pressure data were preprocessed by normalization and feature extraction to obtain processed multi-channel plantar pressure data.

[0007] Furthermore, in one embodiment of the present invention, the head inertial measurement unit is fixed at the center of the wearer's forehead or the center of the back of the head; The trunk inertial measurement unit is fixed to the upper middle section of the wearer's sternum; The eight limb inertial measurement units are respectively fixed to the middle sections of the upper arms of the wearer's two upper limbs, the middle sections of the forearms of the two upper limbs, the front of the middle sections of the two thighs, and the outer sides of the middle sections of the two lower legs; The data collected by the inertial measurement unit includes triaxial acceleration, triaxial angular velocity, and triaxial magnetometer data; Furthermore, in one embodiment of the present invention, the noise suppression is: using complementary filtering or Kalman filtering algorithms to fuse the inertial measurement unit data from each part and eliminate motion noise and environmental interference; The coordinate system one refers to converting the data collected by each inertial measurement unit from the sensor coordinate system to the global coordinate system or the wearer's limb coordinate system; The feature extraction involves calculating the pressure center coordinates, front-to-back and left-to-right pressure distribution ratios, and total pressure value of the flexible pressure sensor array.

[0008] Furthermore, in one embodiment of the present invention, the attitude recognition method further includes: performing static calibration and dynamic calibration on the inertial measurement unit before data acquisition; The static calibration involves the wearer maintaining a standing neutral posture for 3-5 seconds to collect the initial posture reference value; The dynamic calibration involves the wearer performing 3-5 standard shooting or dribbling actions to adapt the system to the range of motion.

[0009] Furthermore, in one embodiment of the present invention, the statistical features include the mean, variance, peak value, energy, and temporal entropy of linear acceleration, angular velocity, attitude angle, and joint angle; The time-domain features include the rise time, fall time, and peak interval of linear acceleration, angular velocity, attitude angle, and joint angle; The features of the multi-channel plantar pressure data include pressure center trajectory, pressure distribution ratio, and ground contact time; The machine learning model includes a lightweight real-time model for real-time recognition of basic actions and a high-level temporal model for recognizing complex and coherent action sequences; wherein the lightweight real-time model is a gradient boosting decision tree, support vector machine or lightweight neural network, and the high-level temporal model is a long short-term memory network, Transformer or 1D-CNN-LSTM hybrid model.

[0010] A posture evaluation method based on wearable devices. The posture evaluation method is based on the basketball action type and its corresponding fused feature vector identified by any posture recognition method described in this invention; the posture evaluation method includes the following steps: Based on the similarity between the fused feature vector and the pre-stored standard action features, a preliminary quality score is output; The identified movements are evaluated in depth using a biomechanical model, which includes preset fusion evaluation indicators and core evaluation content for different types of basketball movements. By using a dynamic time warping algorithm, the wearer's fused feature trajectory is compared with the standard movement feature or personal movement feature baseline through dynamic time warping, which quantifies the degree of movement difference and locates the key nodes of the difference. Based on the results of the quantitative differences in movement and the preset biomechanical model, an injury risk assessment is performed and a personalized assessment report containing specific improvement suggestions is generated.

[0011] Furthermore, in one embodiment of the present invention, the preset fusion evaluation indicators for different basketball action types include: For the shooting motion, the integrated evaluation indicators include trunk pitch angle, upper limb joint angle, head stability, front-to-back pressure distribution ratio of the foot, and trunk and lower limb acceleration at takeoff. For landing cushioning actions, the integrated evaluation indicators include trunk angular velocity and acceleration at the moment of landing, rate of change of lower limb joint angles, trajectory and area of ​​foot pressure center sway, and contact time. For change-of-direction breakthroughs, the integrated evaluation indicators include trunk tilt angle, lower limb joint angle, medial and lateral plantar pressure ratio, and center of gravity movement speed.

[0012] Furthermore, in one embodiment of the present invention, the preset evaluation core content for different basketball action types includes: The core evaluation criteria for shooting motion include measuring body verticality, the continuity of upper and lower limb force, and head stability, and detecting problems such as excessive leaning back, insufficient lower limb force, deformed upper limb movements, and excessive head sway. For landing cushioning, the core assessment includes evaluating trunk stability, lower limb cushioning coordination and overall balance, and predicting the risk of knee, ankle and lower back injuries upon landing. For changes of direction and breakthroughs, the core assessment includes evaluating the efficiency of center of gravity control and identifying risky movements such as excessive trunk tilting and excessive ankle eversion.

[0013] Furthermore, in one embodiment of the present invention, the method further includes establishing a personal motion feature baseline library for the wearer. The personal motion feature baseline library is a fusion feature vector based on multiple high-quality movements of the wearer, and the personal motion feature baseline is established and stored.

[0014] Furthermore, in one embodiment of the present invention, the personalized assessment report of the specific improvement recommendations includes generating specific improvement recommendation texts for force exertion patterns, postural stability, or joint angles based on the results of the injury risk assessment.

[0015] The posture recognition method based on wearable devices described in this invention achieves comprehensive monitoring of athletes' biomechanics and posture information, and can perform objective, real-time, and accurate posture recognition and analysis, providing basketball player training with highly real-time and accurate data. Specific beneficial effects include: 1. The posture recognition method based on wearable devices described in this invention utilizes the deployment and synchronous acquisition of multi-part inertial measurement units and a flexible foot pressure sensor array. After processing such as time synchronization, noise suppression, coordinate system unification, and feature extraction, the spatiotemporal alignment and fusion of multi-source sensor data are achieved. The data obtained by this data fusion method provides an accurate data foundation for the posture recognition of basketball players. Based on this data foundation, accurate and reliable posture recognition of basketball players can be achieved, which can be applied to training guidance systems, training effect evaluation systems, and other aspects of basketball player training.

[0016] 2. Based on the multi-source sensor data obtained by the posture recognition method based on wearable devices described in this invention, the time window of each action is innovatively determined by the sudden change point of the plantar pressure sequence as the event anchor point. By extracting and fusing the statistical features and time domain features of the inertial measurement unit with the features of multi-channel plantar pressure data, a fused feature vector is constructed. The pre-trained machine learning model is then used for recognition, realizing the automatic and accurate recognition of high-dynamic basketball actions such as shooting, landing cushioning, and changing direction breakthrough.

[0017] 3. Based on the posture data obtained by the posture recognition method based on wearable devices described in this invention, a real-time evaluation system for athlete posture can also be researched and developed to automatically evaluate athlete posture. The evaluation results can be used by coaches in formulating training plans.

[0018] 4. The posture assessment method based on wearable devices described in this invention combines a fusion feature vector identifying movement postures with a preset biomechanical model for in-depth assessment, and uses a dynamic time warping algorithm to quantify the difference between the wearer's movements and standard movements or personal movement feature baselines. Finally, it performs injury risk assessment based on the biomechanical model and generates a personalized improvement report, achieving a perfect closed loop from movement recognition to preliminary quality scoring, in-depth assessment, quantification of movement differences, injury risk assessment, and generation of personalized assessment reports.

[0019] 5. The posture recognition and posture evaluation method based on wearable devices described in this invention employs multiple inertial measurement units and two flexible foot pressure sensor arrays, and utilizes a complete set of technologies including multi-source data spatiotemporal alignment, feature fusion, event anchor point window division, machine learning action recognition, dynamic time warping difference quantization, biomechanical model evaluation and risk warning.

[0020] Furthermore, the sensor layout principle of the wearable device described in this invention can be directly mapped to the field of posture perception and recognition technology for humanoid robots. Therefore, based on the posture recognition method described in this invention, functions such as posture perception of each joint of the humanoid robot, force-position coordination control of the end effector, overall motion stability assessment, precise trajectory correction, and closed-loop evaluation of work quality can be achieved. It can also be extended to scenarios such as balance control of bipedal robots and safe interaction of collaborative robots, replacing single vision or encoder solutions. This addresses pain points such as inaccurate posture measurement, unstable control, and poor work consistency in high-dynamic, strong-interference, and occluded environments, achieving a full-link intelligent upgrade from accurate perception to intelligent recognition to real-time evaluation to closed-loop control. This provides quantifiable and replicable core technology support for the research and development and performance optimization of humanoid and bipedal robots.

[0021] The evaluation method obtained by the evaluation system described in this invention can be extended to other sports evaluation fields such as football, volleyball, and tennis, and can also be applied to the fields of posture evaluation, motion control, and human-machine collaboration in intelligent manufacturing for humanoid robots or bipedal robots. Attached Figure Description

[0022] Figure 1 This is a flowchart of the posture recognition method described in Implementation Method 1.

[0023] Figure 2 This is a schematic diagram showing the positions of multiple inertial measurement units on the human body in the method described in Implementation 1.

[0024] Figure 3 This is a flowchart of a basketball posture evaluation method as described in Implementation Method 3.

[0025] Figure 4 The diagram shows the flexible foot pressure sensor array described in the embodiment. In the diagram, 4 represents a pressure sensor unit, and multiple pressure sensor units are arranged in a matrix.

[0026] The attached figures are labeled as follows: Head inertial measurement unit 1, torso inertial measurement unit 2, limb inertial measurement unit 3, and pressure sensor unit 4. Detailed Implementation

[0027] To facilitate understanding of the technical solutions claimed in this invention, the following specific embodiments are provided in conjunction with the accompanying drawings. These embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of this application. Those skilled in the art can make reasonable adjustments to the embodiments given in the following specific embodiments based on their general knowledge in the art to adapt them to specific application scenarios.

[0028] Implementation method one, combined with Figure 1 and Figure 4This embodiment describes a posture recognition method based on a wearable device, which includes the following steps: The wearable device includes multiple inertial measurement units and two flexible foot pressure sensor arrays. The multiple inertial measurement units include a head inertial measurement unit 1 for detecting the wearer's head posture, a torso inertial measurement unit 2 for detecting the wearer's torso posture, and eight limb inertial measurement units 3 for detecting the movement postures of the two upper arms, two forearms, two thighs, and two lower legs, respectively. The posture recognition method is as follows: during the wearer's movement: The attitude data of each inertial measurement unit and the multi-channel plantar pressure data output by each flexible pressure sensor array are collected in real time and synchronously. Using the sequence mutation points of multi-channel plantar pressure data as event anchors, we can detect take-off, landing, sudden stop or turning action events, and further determine the time window of each action event; For each posture data collected within the time window, statistical and temporal features including linear acceleration, angular velocity, posture angle, and joint angle are extracted. At the same time, features of multi-channel plantar pressure data collected synchronously with each posture data are extracted. All statistical features and time-domain features within the aforementioned time window are fused with the features of the multi-channel plantar pressure data to construct a fused feature vector, with each vector corresponding to an action primitive. The fused feature vectors are input into a pre-trained machine learning model to identify basketball action types, including shooting, landing cushioning, and crossover dribbling.

[0029] In this embodiment, after synchronously acquiring the attitude data of each inertial measurement unit and the multi-channel plantar pressure data output by each flexible pressure sensor array in real time, the method further includes multi-source sensor data fusion of the acquired data. This fusion refers to spatiotemporal alignment of all attitude data and multi-channel plantar pressure data acquired at the same time, including: Time alignment: Synchronize all posture data with multi-channel plantar pressure data in time. Spatial alignment of attitude data, noise suppression and coordinate system integration of all attitude data to complete attitude data fusion; All multi-channel plantar pressure data were preprocessed by normalization and feature extraction to obtain processed multi-channel plantar pressure data.

[0030] The posture recognition method based on wearable devices described in this embodiment uses multiple inertial measurement units (IMUs) to collect posture data from different parts of the human body. The positions of the multiple IMUs on the human body can be referenced. Figure 2The positions shown are fixed. The posture data at these positions are sufficient to reflect the movement state of the human torso, head, and limbs. The combination of these posture data can obtain the overall posture of the human body. The wearing positions of all inertial measurement units are, for example, the head inertial measurement unit 1 is fixed at the center of the wearer's forehead or the center of the back of the head.

[0031] The preferred method for fixing the head inertial measurement unit 1 described above ensures the recognition of head posture changes during movement. If placed higher on the head, the sensor would move a greater distance during head movement, resulting in higher sensitivity for recognizing head posture changes. However, this would introduce excessive motion noise. In practice, for head posture changes during basketball, only large pitch and lateral swaying changes need to be monitored. Therefore, placing it at this position ensures sufficient sensitivity without introducing too much motion noise.

[0032] In a preferred embodiment of this invention, the trunk inertial measurement unit 2 is fixed to the upper middle section of the sternum.

[0033] The aforementioned preferred method for fixing the torso inertial measurement unit 2 ensures that this position accurately and stably reflects the overall posture changes of the torso during basketball, especially the pitching, tilting, and twisting movements, providing a reliable spatial reference for calculating the overall posture. In basketball, torso movements often dominate the coordination of the whole body, such as the torso leaning back when shooting, the torso leaning forward and moving laterally when breaking through, and the torso twisting when defending. Fixing the inertial measurement unit to the upper middle part of the sternum can effectively capture the changes in angular velocity and acceleration of the torso in the sagittal, coronal, and horizontal planes, providing core input for subsequent joint angle calculations and force continuity analysis.

[0034] In a preferred embodiment of this invention, the eight limb inertial measurement units 3 are respectively fixed to the middle sections of the upper arms of the wearer's two upper limbs, the middle sections of the forearms of the two upper limbs, the front of the middle sections of the thighs of the two upper limbs, and the outer sides of the middle sections of the lower legs.

[0035] Regarding the fixed positions of the inertial measurement units for the limbs mentioned above: For upper limb posture recognition, two inertial measurement units (IMUs) can be used to identify the posture of the upper arm and forearm respectively to obtain the posture of the entire upper limb. The specific location is in the middle of the upper arm or forearm, preferably in the third of the upper arm near the shoulder, which can be placed near the inner side of the upper arm to avoid impact during movement; the preferred location for the forearm is in the middle of the forearm, between the radius and ulna, avoiding the wrist joint range of motion. This position can directly reflect the flexion and extension of the elbow joint, and the pronation and supination of the forearm, which is especially suitable for monitoring the details of forearm extension and rotation during shooting.

[0036] For lower limb posture recognition, the preferred positions for the two inertial measurement units are the front of the mid-thigh, away from equipment such as trouser pockets and knee pads that may cause friction or obstruction, thus reducing external mechanical interference. The preferred positions for the two inertial measurement units are the outer side of the mid-calf, avoiding the fibular head and ankle joint movement area. This position is less affected by soft tissue vibration during running and jumping, and there is no direct contact with equipment such as shoe uppers and ankle pads, thus avoiding data distortion caused by equipment compression.

[0037] Optionally, inertial measurement units can be deployed 2-3 cm above the wrist joint on both sides of the wearer's wrists to finely analyze wrist rotation during dribbling and the wrist force angle when shooting. An inertial measurement unit was deployed 2-3 cm above the ankle joint on both sides of the wearer to study the ankle posture stability and force direction during crossover dribbling and sudden stop jump shots. Inertial measurement units are deployed in the middle of the instep on both sides of the wearer to capture the foot posture at the moment of takeoff and landing, and to analyze the landing cushioning strategy.

[0038] In a preferred embodiment of this invention, the inertial measurement unit can be implemented using the existing MTi-630AHRS nine-axis sensor. The MTi-630AHRS can measure roll and pitch with an accuracy of no less than ±0.2° and yaw with an accuracy of no less than ±1.5°. It integrates the XKF3 sensor fusion algorithm, which has low static drift characteristics and can support a sampling frequency of up to 1kHz, thereby accurately capturing the details of high-speed movements in basketball.

[0039] The data collected by these inertial measurement units include triaxial acceleration, triaxial angular velocity, and triaxial magnetometer data. Among them, triaxial acceleration can reflect the linear velocity change of the inertial measurement unit's position, triaxial angular velocity can reflect the angular change of the inertial measurement unit's position, and triaxial magnetometer data can reflect the orientation of the inertial measurement unit's position on the horizontal plane.

[0040] Regarding the noise suppression mentioned in this embodiment, a preferred approach is to use complementary filtering or Kalman filtering algorithms to fuse the inertial measurement unit data from various parts and eliminate motion noise and environmental interference.

[0041] The core of denoising by fusing inertial measurement unit data from various parts is to utilize the physical constraints of human motion, namely the strong correlation between the motion of each part, such as the synchronization of arm swing and leg stepping, and the joint angle conforming to the physiological limits of the human body. By combining the complementarity of multi-sensor data, we first perform single-sensor noise reduction, and then use a fusion algorithm to eliminate environmental interference and motion pseudo-noise, ultimately outputting smooth and realistic posture data.

[0042] The coordinate system mentioned in this embodiment is to ensure that the attitude data coordinates of all sensors are consistent, so that the human body's attitude data can be obtained.

[0043] The core of the first coordinate system is to map the position coordinates obtained by all sensors to the global reference coordinate system of the human body, so that the attitude data of all sensors—angle, acceleration, and angular velocity—can be expressed under the same coordinate reference, and realize the fusion calculation of the position information obtained by all sensors to display the overall attitude of the human body.

[0044] The coordinate system follows a holistic principle: first, a global benchmark is established; then, the sensors are bound to the limbs; finally, the coordinate mapping of all data is completed through a transformation matrix; and calibration is performed in conjunction with human kinematic characteristics to eliminate installation or measurement errors and adapt to the dynamic posture changes of the human body during movement.

[0045] A preferred method for unifying the coordinates is to convert the data collected by each inertial measurement unit from the sensor coordinate system to the wearer's limb coordinate system. This method allows for the design of the limb coordinate system based on the wearer's height, limb length, and other body data, as well as the wearing position of each inertial measurement unit.

[0046] Another preferred method for coordinate unification is to transform the data collected by each inertial measurement unit from the sensor coordinate system to the global coordinate system. The data transformed to this coordinate system is more suitable for recognizing human posture and can also recognize changes in human position.

[0047] The coordinate unification can also be achieved using other methods, such as transforming to the coordinate system of the sports field, etc.

[0048] In a preferred embodiment of this invention, the flexible pressure sensor array is a flexible electronic fabric sensing insole.

[0049] The aforementioned flexible electronic fabric sensing insole can be placed directly inside athletic shoes, or between the existing insole and the sole. For example, it can be implemented using the existing matrix-type thin-film pressure sensor RX-M0808MS. In use, it is laid flat on top of the insole inside the athletic shoe, or placed between the insole and the sole. This flexible pressure sensor array consists of a polyester film with excellent comprehensive mechanical properties, highly conductive materials, and nanoscale pressure-sensitive materials. The top layer is a flexible film and a pressure-sensitive layer composited on it, while the bottom layer is a flexible film and conductive circuitry composited on it. The two are bonded together with double-sided adhesive and the sensing area is isolated. When the sensing area is compressed, the disconnected circuitry on the bottom layer becomes conductive through the pressure-sensitive layer on the top layer. The resistance output value at the port changes with the pressure, allowing for real-time and accurate capture of pressure data.

[0050] In a preferred embodiment of this invention, the multi-channel plantar pressure data is acquired using a low-noise, high-gain multi-channel data acquisition device.

[0051] The low-noise, high-gain, multi-channel data acquisition device is a foot-specific data acquisition unit (FPD). The acquisition device and host computer software read and analyze the data. The measurement is performed using the foot-specific data acquisition unit (FPD), and the sensor's measurement area and range can be perfectly adapted to the wearer's foot size.

[0052] The flexible sensor array on the sole of the foot is fixed with breathable adhesive patches to avoid affecting foot movement; the FPD device has a built-in rechargeable lithium battery that can provide power for up to 8 hours and supports up to 64 sensing points.

[0053] In this embodiment, the preferred method for feature extraction is to obtain the pressure center coordinates, front-to-back and left-to-right pressure distribution ratios, and total pressure value of the flexible pressure sensor array using multi-channel plantar pressure data. First, the data collected by the multi-source sensors are preprocessed to remove invalid values, reduce noise through combined filtering, and confirm the consistency between the sensor coordinate calibration and the local coordinate system of the foot.

[0054] Based on the effective sampling points and their pressure areas, the total pressure value, the front-to-back or left-to-right pressure distribution ratio (dividing the corresponding areas of the sole of the foot to calculate the pressure ratio), and the pressure center coordinates (weighted pressure center) are calculated in sequence. Finally, the accuracy is optimized through pressure value, coordinate calibration, and parameter logic consistency verification, and the output is stored in a standardized format.

[0055] The above data processing procedure takes into account both data accuracy and engineering practicality throughout, and is suitable for scenarios such as gait analysis.

[0056] The data acquisition described in this embodiment is synchronous acquisition. In order to further ensure the accuracy of the data, the time information of all data is adjusted by time alignment to achieve data alignment on the time axis, thereby ensuring that the data obtained by all sensors at the same moment can truly reflect the posture of the human body wearing these sensors at that moment.

[0057] Implementation Method 2 is a further optimization of the posture recognition method based on wearable devices described in Implementation Method 1. In this implementation method, the posture recognition method further includes: performing static calibration and dynamic calibration on the inertial measurement unit before data acquisition.

[0058] In this embodiment, static calibration requires a series of continuous actions: (1) The wearer stands upright, looks straight ahead, and holds the neutral posture with his hands hanging down naturally for 2 seconds; (2) Change to having both arms outstretched, holding the ball above the head, and hold for 2 seconds; (3) Change to rotating your body to the left as far as possible while holding the ball in front of your chest with both hands for 2 seconds; (4) Finally, rotate your body to the right as far as possible while holding the ball in front of your chest with both hands for 2 seconds.

[0059] After the above actions are collected, the sensor data is processed to determine the initial and limit values.

[0060] In this embodiment, a preferred method for dynamic calibration is provided, which involves the wearer performing natural walking and running movements, defensive actions, two-handed chest passes, and 3-5 shooting and dribbling actions. During each action, data is continuously collected at a frequency higher than 2kHz. Based on this data, the range of variation of sensor data collected at various positions during the shooting motion is obtained. These ranges will help us eliminate abnormal data.

[0061] In this embodiment, the preferred method for detecting the time window of take-off, landing, sudden stop or turning action events is to first locate the key anchor point of the action through a unified mutation point judgment rule, then define the start and end boundaries of the exclusive time window based on the biomechanical model of each action, and finally combine the temporal and spatial distribution characteristics of the plantar pressure within the window to complete the accurate classification of the action type, rather than directly judging the action with a single general window.

[0062] A mutation point is a critical time node in plantar pressure data that transitions from steady state to transient state, or vice versa. It reflects sudden changes in plantar force, such as sudden increases, decreases, or reconfigurations in pressure distribution. It serves as the core anchor point for all motion events. A dual criterion of "single-channel feature + multi-channel fusion" is used; any time point that satisfies any mutation type and passes significance verification is considered a valid mutation point. Time sequence for each plantar pressure channel Where i is the channel number and t is the timestamp, calculate the first-order difference. and second-order difference Determine the mutation type by combining differential features: The pressure value increased or decreased sharply within 13 sampling points, and the absolute value of the first difference exceeded the threshold. For a step-type mutation, the threshold The normalized value is 0.2-0.3. For example, the sudden increase in heel channel pressure from 0.1 to 0.8 before landing and the sudden decrease in forefoot channel pressure from 0.9 to 0.1 after takeoff is a step-like abrupt change. The slope of the pressure change suddenly reverses or increases sharply, and the absolute value of the second difference exceeds the threshold. The The normalized value is 0.15-0.25, reflecting abrupt changes in acceleration due to force changes, which are slope-dependent abrupt changes. For example, during an emergency stop, the pressure in the lateral forefoot channel changes from a slow increase to a rapid surge, and the second-order difference significantly deviates from the steady state.

[0063] The core mutation characteristic of the pressure value changing from 0 to non-zero, or from non-zero to 0, and remaining greater than or equal to 2 sampling points, is the foot contact or off-ground pressure value, which belongs to zero-crossover mutation.

[0064] A sudden change in a single channel may be a localized change in force, requiring verification through the spatiotemporal correlation of multiple channel sudden changes. The event anchor point is considered a valid event anchor point if it meets any of the following conditions: At least two related pathways, such as the medial and lateral heel or the medial and lateral forefoot, undergo the above mutation within the same time window, which is called synchronous mutation; The abrupt changes in the associated channels in a biomechanical sequence, such as the heel channel changing first upon landing and the forefoot channel changing first, or the forefoot channel changing first upon takeoff and the heel channel changing first, with a maximum time interval of 5 sampling points, are considered temporal abrupt changes. Total plantar pressure: Where n is the number of channels and j is the channel number. Let be the pressure value of the j-th channel at time t. If a step-type abrupt change occurs and the absolute value change is not less than 0.3 after normalization, it reflects a sudden change in the overall force, which is called the total pressure abrupt change.

[0065] Optionally, to further exclude minor, non-action-related mutations, the mutation intensity at the mutation point can be calculated. , The standard deviation of total pressure, As the first difference of the total pressure, set ≥3 is considered a significant mutation point (the 3σ principle for steady-state data), and only significant mutation points are retained as event anchors.

[0066] The method for determining the start and end points of the time window can be based on biomechanical principles. For example, the core biomechanical characteristics of takeoff are: before takeoff, the sole of the foot is fully in contact with the ground and the force is stable; during takeoff, the forefoot exerts force first, at which point the pressure increases sharply; subsequently, the heel leaves the ground, causing the pressure to drop sharply; finally, the entire sole leaves the ground, bringing the total pressure to zero. The core anchor point... The starting moment of the step-like sudden increase in the forefoot-related channel is the moment when a person jumps and exerts force.

[0067] Based on the above principles, the method for determining the start and end points of the time window is illustrated with an example: Window start point :from The system shifts forward until the first time point when the total plantar pressure reaches a steady state, which is the steady-state determination point. At this point, the coefficient of variation (CV) of the total pressure is no higher than 5% and remains at least 3 sampling points. If the shift is no higher than 10 sampling points, the steady state is reached, and the shift endpoint is taken as the starting point of the window. If the offset is greater than 10 sampling points, take... 10 is the starting point of the window. .

[0068] Window end point :from The time point at which the total plantar pressure first reaches zero and remains at least 2 sampling points after shifting backward represents the moment the entire plantar surface leaves the ground; if the shift reaches the departure point at no more than 20 sampling points, the end point of the shift is taken as the end point of the window. If the offset is greater than 20 sampling points, take... +20 is the end point of the window. .

[0069] Window constraints: the above window end point and window start point The area between these points is the window constraint, which contains no more than 30 sampling points. These sampling points cover the core process from takeoff to liftoff.

[0070] The statistical features described in this embodiment include the mean, variance, peak value, energy, and temporal entropy of linear acceleration, angular velocity, attitude angle, and joint angle.

[0071] The time-domain features described in this embodiment include linear acceleration, angular velocity, attitude angle, and the rise time, fall time, and peak interval of joint angles.

[0072] The features in the multi-channel plantar pressure data described in this embodiment include the pressure center trajectory, pressure distribution ratio, and ground contact time.

[0073] The machine learning model described in this embodiment includes a lightweight real-time model for real-time recognition of basic actions and a high-level temporal model for recognizing complex and coherent action sequences; wherein, the lightweight real-time model is a gradient boosting decision tree, a support vector machine, or a lightweight neural network, and the high-level temporal model is a long short-term memory network, a Transformer, or a 1D-CNN-LSTM hybrid model. The basketball action types described in this embodiment include shooting, landing cushioning, and crossover dribbling.

[0074] Implementation method three, see Figure 3This embodiment describes a posture evaluation method based on the basketball action type and its corresponding fused feature vector identified by the posture recognition method described in Embodiment 1 or 2. The posture evaluation method includes the following steps: Obtain the fused feature vector of the athlete's reaction movement posture during the movement process; Based on the similarity between the fused feature vector and the pre-stored standard action features, a preliminary quality score is output; The identified movements are evaluated in depth using a biomechanical model, which includes preset fusion evaluation indicators and core evaluation content for different types of basketball movements. By using a dynamic time warping algorithm, the wearer's fused feature trajectory is compared with the standard movement feature or personal movement feature baseline through dynamic time warping, which quantifies the degree of movement difference and locates the key nodes of the difference. Based on the results of the quantitative differences in movement and the preset biomechanical model, an injury risk assessment is performed and a personalized assessment report containing specific improvement suggestions is generated.

[0075] This embodiment provides a preferred method for the preset fusion evaluation index for different basketball action types, including: For the shooting motion, the integrated evaluation indicators include trunk pitch angle, upper limb joint angle, head stability, front-to-back pressure distribution ratio of the foot, and trunk and lower limb acceleration at takeoff. For landing cushioning actions, the integrated evaluation indicators include trunk angular velocity and acceleration at the moment of landing, rate of change of lower limb joint angles, trajectory and area of ​​foot pressure center sway, and contact time. For change-of-direction breakthroughs, the integrated evaluation indicators include trunk tilt angle, lower limb joint angle, medial and lateral plantar pressure ratio, and center of gravity movement speed.

[0076] The evaluation criteria for whether a change-of-direction breakthrough move is qualified are shown in the table below: Table 1

[0077] As described in Table 1, this implementation method pre-sets four evaluation indicators and their threshold ranges for the change of direction breakthrough action, and judges whether the action is qualified or defective based on whether the measured value of each indicator meets the threshold conditions and the action performance.

[0078] If the trunk tilt angle is within the range of 8° to 20° and the tilt direction is consistent with the change direction, the movement is considered qualified. If it is less than 8°, it is considered insufficient tilt. If it is greater than 20°, it is considered excessive tilt and body imbalance.

[0079] Thresholds are set for the knee and ankle joint angles in the lower limb joint angles. If the knee joint angle is within the range of 100° to 150° and the ankle joint angle is within the range of 75° to 100°, and the angle fluctuation during the force exertion is no more than 15°, the movement is evaluated as qualified. If the knee or ankle joint angle exceeds the threshold, or the angle fluctuation during the force exertion is greater than 15°, it is considered to be in the state between take-off and landing, which is not a change of direction breakthrough movement.

[0080] Thresholds are set for the pressure ratios on the inner and outer sides of the foot for inner and outer lateral changes of direction. If the inner lateral change of direction is between 1.3 and 3.0 or the outer lateral change of direction is between 0.3 and 0.7, the movement is considered acceptable. If the pressure ratio is between 0.7 and 1.3, the movement is considered defective, which may be due to walking in a straight line or turning.

[0081] The threshold for center of gravity movement speed is 1.2m / s to 3.5m / s. If the center of gravity movement speed remains within the threshold during the movement and the speed direction deviation is not less than 30°, the movement is considered qualified. If it exceeds the threshold, the movement is considered defective. If the center of gravity movement speed is less than 1.2m / s, the speed is too slow. If the center of gravity movement speed is greater than 3.5m / s, the speed is too fast and may be a straight sprint.

[0082] The torso tilt angle is the angle between the torso and the ground normal.

[0083] The lower limb in the lower limb joint angle refers to the supporting leg, and the knee joint angle and ankle joint angle refer to the angle at the moment the foot lands.

[0084] The shooting motion and landing cushioning motion can be evaluated in a similar manner, and will not be exemplified in this implementation.

[0085] This embodiment provides a preferred method for the preset evaluation core content for different basketball action types, including: The core evaluation criteria for shooting motion include measuring body verticality, the continuity of upper and lower limb force, and head stability, and detecting problems such as excessive leaning back, insufficient lower limb force, deformed upper limb movements, and excessive head sway. For landing cushioning, the core assessment includes evaluating trunk stability, lower limb cushioning coordination and overall balance, and predicting the risk of knee, ankle and lower back injuries upon landing. For changes of direction and breakthroughs, the core assessment includes evaluating the speed of the center of gravity shift and identifying any risky movements such as excessive trunk tilt or excessive ankle eversion.

[0086] This embodiment proposes a preferred method for the personalized evaluation report of the specific improvement suggestions, which includes generating specific improvement suggestion texts for force application patterns, posture stability, or joint angles based on the results of the injury risk assessment.

[0087] For example, regarding the core issue of Zhang's right-side change of direction breakthrough, the following evaluation results are shown in Table 2: Table 2

[0088] Based on the above evaluation results, the following improvement recommendations are generated: Mr. Zhang's trunk tilt angle is 6°, which is below the lower limit of the standard. This results in insufficient shift of the body's center of gravity during changes of direction, leading to poor force transmission and a medium risk of insufficient speed and imbalance during changes of direction. It is recommended that during training, he consciously tilt his trunk to the right during right-side changes of direction. Assistive devices such as balance mats can be used for tilt stability training. Each training session should consist of 10 sets, with each set maintaining a trunk tilt angle close to the lower limit of the standard. Within the range, each set lasts 15 seconds, improving trunk tilt control.

[0089] The supporting leg's knee angle is 105° and ankle angle is 70°, slightly below the standard threshold, and is considered low-risk. This results in insufficient cushioning in the lower limbs during force exertion, hindering the transmission of explosive power. It is recommended that during training, the knee bend to 110° at the moment of force exertion during a change of direction. 120°, ankle joint maintained at 75° 80° can be improved by lightly weighted half squats and calf raises to enhance the precision of lower limb joint angle control. Each training session consists of 15 sets of 10 repetitions, with a focus on feeling the stability of the joint angle at the moment of exertion.

[0090] The center of gravity shift speed of 1.0 m / s is slightly lower than the standard, which is related to insufficient trunk tilt and inadequate cushioning of the lower limb joints. It is recommended to optimize the sequence of force exertion during directional changes, first guiding the center of gravity shift by tilting the trunk to the side, and then using the cushioning force of the supporting leg's knee and ankle joints to propel the center of gravity to move quickly; linear acceleration combined with directional change training can be performed, 10 repetitions per set, 3 sets per training session, gradually increasing the center of gravity shift speed during directional changes to above 1.2 m / s.

[0091] The core evaluation content of shooting motion and landing cushioning motion can be evaluated in a similar way to generate personalized evaluation reports, and will not be exemplified in this implementation.

[0092] The posture evaluation method described in this embodiment may further include establishing a personal motion feature baseline library for the wearer. The personal motion feature baseline library is a fusion feature vector based on multiple high-quality movements of the wearer, and the personal motion feature baseline is established and stored. The data structure of the individual motion feature baseline database is shown in the table below: The Athlete Basic Information Table stores the basic personal information of athletes and serves as the main table of the baseline database. It is associated with other data tables, as shown in Table 3.

[0093] Table 3

[0094] The athlete's unique identifier ID is a custom code; The action type table stores the types of actions that basketball players need to evaluate, such as crossover dribble, jump shot, and stop pass. It is associated with the action feature data table, as shown in Table 4.

[0095] Table 4:

[0096] The motion feature baseline table stores the core feature baseline values ​​of individual athletes and individual motions. It is the core reference for quantitative difference calculation and is associated with the athlete table and motion type table, as shown in Table 5.

[0097] Table 5:

[0098] The baseline data collection number is no less than 5 standard actions.

[0099] The sensor acquisition parameter table is used to store the sensor parameters during baseline data acquisition, ensuring consistent acquisition conditions during subsequent evaluations and improving the accuracy of baseline references, as shown in Table 6.

[0100] Table 6:

[0101] The baseline update log table is used to record the update records of the baseline library, which makes it easy to trace baseline changes, such as updating the baseline value after the athlete's movement improves, and linking the athlete table and the baseline table, as shown in Table 7.

[0102] Table 7:

[0103] The update refers to the clearly modified baseline indicators and values; the operator is an evaluation coach or data analyst.

[0104] During the assessment, the athlete's basic information is first retrieved by athlete ID, then the corresponding action feature baseline value is matched according to the assessment action ID, and combined with the sensor-collected parameters, the quantitative difference between the actual measurement value and the baseline value is calculated, and finally an assessment report is generated. During training or testing, when high-risk movements are detected, such as a tendency for the knee joint to buckle inward upon landing, or when the thigh posture calculated by the trunk inertial measurement unit is combined with the joint angle obtained by the lower limb inertial measurement unit and the plantar eversion pressure pattern, a warning sound is issued in real time via wireless headphones or field-side equipment. Secondly, a visualization report is generated, producing a "3D movement comparison report," which reconstructs the wearer's fused data into a simplified 3D model of the whole body and compares it with the standard model on the same screen, intuitively showing the differences and quantitative deviations of multiple parts, such as a 5-degree deviation of the take-off angle to the left, a 10-degree deviation of the upper limb joint angle, and uneven distribution of plantar pressure.

[0105] Implementation Method Four: This implementation method describes the application of the posture recognition method described in Implementation Method One or Two, and the posture evaluation method described in Implementation Method Three. Specifically, the inertial measurement unit is deployed in each joint or link of the industrial robot, and the flexible pressure sensor array is deployed in the robot's end effector.

[0106] By implementing the posture recognition method described in Embodiment 1 or 2, the posture of an industrial robot can be recognized.

[0107] The posture evaluation method described in Implementation Method 3 can achieve accurate identification of robot motion trajectory, stability evaluation of joint posture, and closed-loop optimization of end-effector force control through data acquisition, spatiotemporal alignment, event anchor point detection, feature fusion, and machine learning recognition steps.

[0108] This embodiment is intended to illustrate that any of the technical solutions claimed in this invention can be directly transferred to the field of intelligent manufacturing to improve the accuracy, efficiency and safety of intelligent manufacturing.

Claims

1. A posture recognition method based on wearable devices, characterized in that, The wearable device includes multiple inertial measurement units and two flexible foot pressure sensor arrays; the multiple inertial measurement units include a head inertial measurement unit (1) for detecting the wearer's head posture, a trunk inertial measurement unit (2) for detecting the wearer's torso posture, and eight limb inertial measurement units (3) for detecting the movement postures of the two upper arms, the two upper forearms, the two thighs, and the two lower legs, respectively. The posture recognition method is as follows: during the wearer's movement: The attitude data of each inertial measurement unit and the multi-channel plantar pressure data output by each flexible pressure sensor array are collected in real time and synchronously. Using the sequence mutation points of multi-channel plantar pressure data as event anchors, we can detect take-off, landing, sudden stop or turning action events, and further determine the time window of each action event; For each posture data collected within the time window, statistical and temporal features including linear acceleration, angular velocity, posture angle, and joint angle are extracted. At the same time, features of multi-channel plantar pressure data collected synchronously with each posture data are extracted. All statistical features and time-domain features within the aforementioned time window are fused with the features of the multi-channel plantar pressure data to construct a fused feature vector, with each vector corresponding to an action primitive. The fused feature vectors are input into a pre-trained machine learning model to identify basketball action types, including shooting, landing cushioning, and crossover dribbling.

2. The posture recognition method based on a wearable device according to claim 1, characterized in that, After real-time synchronous acquisition of attitude data from each inertial measurement unit and multi-channel plantar pressure data output from each flexible pressure sensor array, the process also includes multi-source sensor data fusion. This fusion refers to spatiotemporal alignment of all attitude data and multi-channel plantar pressure data acquired simultaneously, including: Time alignment: Synchronize all posture data with multi-channel plantar pressure data in time. Spatial alignment of attitude data, noise suppression and coordinate system integration of all attitude data to complete attitude data fusion; All multi-channel plantar pressure data were preprocessed by normalization and feature extraction to obtain processed multi-channel plantar pressure data.

3. The posture recognition method based on a wearable device according to claim 2, characterized in that, The noise suppression is achieved by using complementary filtering or Kalman filtering algorithms to fuse inertial measurement unit data from various parts and eliminate motion noise and environmental interference. The coordinate system one refers to converting the data collected by each inertial measurement unit from the sensor coordinate system to the global coordinate system or the wearer's limb coordinate system; The feature extraction involves calculating the pressure center coordinates, front-to-back and left-to-right pressure distribution ratios, and total pressure value of the flexible pressure sensor array.

4. A posture recognition method based on a wearable device according to any one of claims 1-3, characterized in that, The attitude recognition method further includes performing static and dynamic calibration on the inertial measurement unit before data acquisition; The static calibration involves the wearer maintaining a standing neutral posture for 3-5 seconds to collect the initial posture reference value; The dynamic calibration involves the wearer performing 3-5 standard shooting or dribbling actions to adapt the system to the range of motion.

5. The posture recognition method based on a wearable device according to claim 1, characterized in that, The statistical characteristics include the mean, variance, peak value, energy, and temporal entropy of linear acceleration, angular velocity, attitude angle, and joint angle; The time-domain features include the rise time, fall time, and peak interval of linear acceleration, angular velocity, attitude angle, and joint angle; The features of the multi-channel plantar pressure data include pressure center trajectory, pressure distribution ratio, and ground contact time; The machine learning model includes a lightweight real-time model for real-time recognition of basic actions and a high-level temporal model for recognizing complex and coherent action sequences; wherein the lightweight real-time model is a gradient boosting decision tree, support vector machine or lightweight neural network, and the high-level temporal model is a long short-term memory network, Transformer or 1D-CNN-LSTM hybrid model.

6. A posture evaluation method based on wearable devices, characterized in that, The posture evaluation method is based on the basketball action type and its corresponding fused feature vector identified by the posture recognition method according to any one of claims 1-5; the posture evaluation method includes the following steps: Based on the similarity between the fused feature vector and the pre-stored standard action features, a preliminary quality score is output; The identified movements are evaluated in depth using a biomechanical model, which includes preset fusion evaluation indicators and core evaluation content for different types of basketball movements. By using a dynamic time warping algorithm, the wearer's fused feature trajectory is compared with the standard movement feature or personal movement feature baseline through dynamic time warping, which quantifies the degree of movement difference and locates the key nodes of the difference. Based on the results of the quantitative differences in movement and the preset biomechanical model, an injury risk assessment is performed and a personalized assessment report containing specific improvement suggestions is generated.

7. The posture evaluation method based on a wearable device according to claim 6, characterized in that, The preset fusion evaluation indicators for different basketball movement types include: For the shooting motion, the integrated evaluation indicators include trunk pitch angle, upper limb joint angle, head stability, front-to-back pressure distribution ratio of the foot, and trunk and lower limb acceleration at takeoff. For landing cushioning actions, the integrated evaluation indicators include trunk angular velocity and acceleration at the moment of landing, rate of change of lower limb joint angles, trajectory and area of ​​foot pressure center sway, and contact time. For change-of-direction breakthroughs, the integrated evaluation indicators include trunk tilt angle, lower limb joint angle, medial and lateral plantar pressure ratio, and center of gravity movement speed.

8. The posture evaluation method based on a wearable device according to claim 6, characterized in that, The core evaluation content preset for different types of basketball movements includes: The core evaluation criteria for shooting motion include measuring body verticality, the continuity of upper and lower limb force, and head stability, and detecting problems such as excessive leaning back, insufficient lower limb force, deformed upper limb movements, and excessive head sway. For landing cushioning, the core assessment includes evaluating trunk stability, lower limb cushioning coordination and overall balance, and predicting the risk of knee, ankle and lower back injuries upon landing. For changes of direction and breakthroughs, the core assessment includes evaluating the efficiency of center of gravity control and identifying risky movements such as excessive trunk tilting and excessive ankle eversion.

9. The posture evaluation method based on a wearable device according to claim 6, characterized in that, The method also includes establishing a personal motion feature baseline library for the wearer, which is established and stored based on the fused feature vectors of multiple high-quality movements of the wearer.

10. The posture evaluation method based on a wearable device according to claim 6, characterized in that, The personalized assessment report with specific improvement recommendations includes textual recommendations for specific improvements to force patterns, postural stability, or joint angles, based on the results of the injury risk assessment.

Images