CN117549306A - Fall position identification method and accompanying robot - Google Patents

Abstract

The invention discloses a falling position identification method and a companion robot, which can quickly identify falling behaviors of VR experimenters by early warning the falling behaviors through human body posture angles and angular velocities acquired by VR head-mounted equipment, then predict human head track based on an LSTM coding and decoding model, and have high model prediction accuracy, and can be used for path planning of the companion robot.

Description

A fall position recognition method and companion robot

Technical field

The invention relates to the field of fall posture recognition, in particular to a fall position recognition method and a companion robot.

Background technique

VR experiencers are prone to falls due to various reasons in daily life, such as encountering uneven ground or obstacles when using VR equipment, and the virtual environment does not show the smoothness of the obstacles and the actual ground, which leads to Falling; or due to poor quality or stability of the VR equipment, which may cause the experiencer to lose balance and fall when moving or turning; or due to the VR experience, the experiencer may need to move or make quick movements, if the experiencer Poor physical balance or insufficient perception of the virtual environment may lead to falls.

If no one discovers and helps in time, it may cause serious consequences. It was reported that a 44-year-old Moscow resident unfortunately fell on a glass table while using a VR headset and eventually suffered scratches. Bleed to death. Although existing monitoring systems or smart devices can identify falls to a certain extent, they cannot accurately determine the state of the person who fell, and cannot protect people and VR machines after a fall. Therefore, it is of great significance to develop a system that can accurately identify falls and automatically pick up people for rescue.

Contents of the invention

The purpose of the present invention is to provide a fall position recognition method and a companion robot, which can correctly and quickly identify the VR experiencer's falling behavior and the landing position of the head, and protect the VR experiencer by traveling to the landing position of the head.

A fall location identification method includes the following steps:

S1. Train LSTM neural network;

Data collection: Collect a time series of position coordinate data of the VR head-mounted device. The position coordinate data of each time step in the time series includes the coordinates (xi, yi, zi), i of the VR head-mounted device in the navigation coordinate system. is the number of the time step, x is the coordinate in the east direction, y is the coordinate in the north direction, and z is the coordinate perpendicular to the horizontal plane;

Data preparation: Standardize the values corresponding to the three coordinate axes of the navigation coordinate system for the position coordinate data of each time step in the time series, so that the position coordinate data of each time step in the standardized time series can be used in the navigation system. The mean value of the values corresponding to the three coordinate axes of the coordinate system is 0 and the variance is 1;

Divide the data set: Divide the standardized time series into input sequence and target sequence;

LSTM neural network training: Initialize the LSTM model parameters, perform forward propagation of the input sequence through the LSTM model to obtain the prediction sequence, calculate the loss function based on the prediction sequence and the target sequence, and update the LSTM according to the gradient of the calculated loss function through the back propagation algorithm. The weight parameters of the model, repeat the training process until the preset training times are reached or until convergence;

S2. Using the ground as a reference, establish a navigation coordinate system in the VR experience space, and collect real-time data of the VR head-mounted device when the human body wears the VR head-mounted device, including attitude angle, acceleration and position coordinate data in the navigation coordinate system. ;

S3. Input attitude angle and acceleration into the fall prediction model and extract real-time data that triggers fall conditions;

S4. Use the sliding window method to extract the position coordinate data corresponding to the time steps within the sliding window range from the real-time data extracted in step S3, forming a sample time series. The time steps within the sliding window range include the real-time data extracted in S3. The time step of the data;

S5. Standardize the S4 sample time series and input it into the trained LSTM neural network. The output result is denormalized to obtain the prediction result of the position coordinate data of the VR head-mounted device.

The steps of the standardization process specifically include:

Calculate the mean and variance of the values corresponding to the three coordinate axes of the navigation coordinate system for the position coordinate data of all time steps;

Use the following formula to convert the position coordinate data of each time step in the time series into standardized values from the values corresponding to the three coordinate axes of the navigation coordinate system:

Standardized value = (original value - mean) / variance;

The original value is the value corresponding to the selected coordinate axis in the navigation coordinate system of the position coordinate data of each time step. The mean value is the mean value and the variance of the value corresponding to the selected coordinate axis of the position coordinate data of all time steps in the navigation coordinate system. It is the variance of the value corresponding to the selected coordinate axis of the navigation coordinate system for the position coordinate data of all time steps.

Preferably, when dividing the data set, the sliding window method is used, the standardized time series within a time window is used as the input sequence, and the time series of the next time step of the input sequence is used as the target sequence.

Furthermore, when dividing the data set, the sliding window method is used to divide the standardized time series within a time window into time series of three consecutive time steps. The time series of the first two time steps are used as the input sequence, and the last time series is used as the input sequence. The time series of steps is used as the target sequence.

In step S2, the attitude angle includes pitch angle, roll angle, and yaw angle. The yaw angle represents the offset angle from the Y-axis; the pitch angle represents the offset angle from the X-axis; the roll angle represents the deflection angle from the Z-axis, and the X-axis is The axis is perpendicular to the left and right sides of the human body, the Z axis is the axis perpendicular to the front and rear sides of the human body, the Y axis is perpendicular to the X axis and Z axis, and the acceleration is the acceleration AccX, AccY, and AccZ along the X, Y, and Z axes.

The fall prediction model in step S3 includes an attitude angle distinction unit and an acceleration distinction unit. The fall prediction model performs the following steps to determine whether a fall condition is triggered:

After the posture angle is input into the posture angle distinction unit, formula (1) can be used to distinguish the forward bending and backward falling of the human body from other postures, and obtain sqrtRP;

sqrtRP calculates RPP_delta according to formula (2), and compares it with the set RPP_delta threshold, which can effectively extract the sideways and backward bending and falling movements of the human body.

RPP_delta=fabs(SqrtRP-yaw) (2)

When the RPP_delta angle is greater than 30°, the acceleration is input into the acceleration differentiation unit;

When the yaw angle drops to no more than 70°, the acceleration differentiation unit first calculates the combined acceleration of the Z-axis and the Y-axis according to formula (3) based on the acceleration in the Z-axis direction and the Y-axis direction: SqrtAccYZ, and then based on SqrtAccYZ, Z The acceleration in the axial direction and formula (4) are used to calculate ACCYZZ_delta, and then compared with the normal threshold and fall threshold to determine whether the human body has fallen;

AccYZZ_delta=fabs(SqrtAccYZ-AccZ) (4)

Observe the change of ACCYZZ_delta angle. When ACCYZZ_delta is less than 40, it is judged as a fall and the fall condition is triggered.

A kind of companion robot goes to the position coordinates predicted in step S5 for protection. The companion robot carries a memory foam pad to absorb impact and energy and reduce the impact of the head when falling.

The present invention provides a human head trajectory prediction method based on the LSTM encoding and decoding model. The model prediction accuracy is high and can be used for path planning of accompanying robots. The present invention also provides an early warning method for falling behavior, which can quickly identify VR The experiencer’s falling behavior.

Description of the drawings

Figure 1 is a schematic diagram of the LSTM neural network;

Figure 2 is a schematic diagram of the posture angle of the human body;

Figure 3 is a schematic diagram of the fall judgment process of the fall prediction model;

Figure 4 is a flow chart of companion robot path planning.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention. That is, the described embodiments are only some embodiments of the present invention, rather than all embodiments. The components of the embodiments of the invention generally described and represented in the figures herein may be arranged and designed in a variety of different configurations.

Therefore, the following detailed description of the embodiments of the invention provided in the appended drawings is not intended to limit the scope of the claimed invention, but rather to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without any creative work fall within the scope of protection of the present invention.

It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that these entities or operations are mutually exclusive. any such actual relationship or sequence exists between them. Furthermore, the terms "comprises," "comprises," or any other variations thereof are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that includes a list of elements includes not only those elements, but also those not expressly listed other elements, or elements inherent to the process, method, article or equipment. Without further limitation, an element defined by the statement "comprises a..." does not exclude the presence of additional identical elements in a process, method, article, or apparatus that includes the stated element.

The features and performance of the present invention will be described in further detail below with reference to examples.

Example 1

This application proposes a fall location identification method, the details are as follows:

S1. Training LSTM neural network

Data collection: Collect a time series of position coordinate data of the VR head-mounted device. The position coordinate data of each time step in the time series includes the coordinates (xi, yi, zi), i of the VR head-mounted device in the navigation coordinate system. is the number of the time step, x is the coordinate in the east direction, y is the coordinate in the north direction, and z is the coordinate perpendicular to the horizontal plane;

Data preparation: Standardize the values corresponding to the three coordinate axes of the navigation coordinate system for the position coordinate data of each time step in the time series, so that the position coordinate data of each time step in the standardized time series can be used in the navigation system. The mean value of the values corresponding to the three coordinate axes of the coordinate system is 0 and the variance is 1;

Divide the data set: Divide the standardized time series into input sequence and target sequence;

LSTM neural network training: Initialize the LSTM model parameters, perform forward propagation of the input sequence through the LSTM model to obtain the prediction sequence, calculate the loss function based on the prediction sequence and the target sequence, and update the LSTM according to the gradient of the calculated loss function through the back propagation algorithm. weight parameters of the model, and the training process is repeated until the preset training times are reached or until convergence.

S2. Using the ground as a reference, establish a navigation coordinate system in the VR experience space, and collect real-time data of the VR head-mounted device when the human body wears the VR head-mounted device, including attitude angle, acceleration and position coordinate data in the navigation coordinate system. ;

S3. Input attitude angle and acceleration into the fall prediction model and extract real-time data that triggers fall conditions;

S4. Use the sliding window method to extract the position coordinate data corresponding to the time steps within the sliding window range from the real-time data extracted in step S3, forming a sample time series. The time steps within the sliding window range include the real-time data extracted in S3. The time step of the data;

S5. Standardize the S4 sample time series and input it into the trained LSTM neural network, and perform inverse normalization on the output results to obtain the prediction results of the position coordinate data of the VR head-mounted device;

Specifically, the position coordinate data of the VR head-mounted device is obtained by installing a radar sensor on the VR head-mounted device. The antenna in the radar sensor emits electromagnetic waves. The emitted electromagnetic waves propagate in space and will be reflected when encountering objects. Refraction or scattering. The receiver in the radar sensor receives the signal reflected by the target, and the received signal undergoes processing steps such as amplification, filtering, and modulation to extract relevant information about the target. By measuring the time difference between transmitting and receiving signals, radar sensors can calculate the distance to a target.

Typically, radar sensors use a navigation coordinate system referenced to the local ground, and objects in the VR experience space have defined positions relative to this coordinate system. Therefore, the position coordinates of the VR head-mounted device in the VR experience space can be obtained: (xi, yi, zi).

Preferably, the steps of standardization include:

Calculate the mean and variance of the values corresponding to the three coordinate axes of the navigation coordinate system for the position coordinate data of all time steps;

Use the following formula to convert the position coordinate data of each time step in the time series into standardized values from the values corresponding to the three coordinate axes of the navigation coordinate system:

Standardized value = (original value - mean) / variance;

The original value is the value corresponding to the selected coordinate axis in the navigation coordinate system of the position coordinate data of each time step. The mean value is the mean value and the variance of the value corresponding to the selected coordinate axis of the position coordinate data of all time steps in the navigation coordinate system. It is the variance of the value corresponding to the selected coordinate axis of the navigation coordinate system for the position coordinate data of all time steps.

Through the above standardization process, the mean value of the standardized value is 0 and the variance is 1.

Inverse normalization processing is used to calculate the true value based on the obtained standardized prediction results.

True value = standardized prediction result * variance + mean

In step S1 of this embodiment, the input sequence is a standardized historical time series used for model input, and the target sequence is a standardized time series to be predicted at the next time step.

Specifically, using the sliding window method, the standardized time series within a time window is used as the input sequence, and the time series of the next time step of the input sequence is used as the target sequence.

Furthermore, when using the sliding window method, the standardized time series within a time window is divided into time series of three consecutive time steps. The time series of the first two time steps are used as the input sequence, and the time series of the last time step is used as the input sequence. as the target sequence.

Assume that the time series after normalization is as follows:

t1:10,20,30,40;

t2:15,25,35,45;

t3:20,30,40,50;

t4:25,35,45,55;

Use the sliding window method to convert these data into input and target sequences:

Assume that the window size is 2, that is, each input sequence sample contains data for 2 time steps:

Input:[10,20,30,40],[15,25,35,45], output:[20,30,40,50]

Input:[15,25,35,45],[20,30,40,50], output:[25,35,45,55]

The sliding window method can divide the continuous VR head-mounted device position data into input windows and corresponding outputs.

In step S1, the LSTM neural network is a special RNN (recurrent neural network). As shown in Figure 1, the LSTM network structure inputs the position coordinates of the VR head-mounted device at time t into the network in chronological order. In the LSTM unit It is necessary to calculate the output of the current state based on the state of the memory unit (cell) combined with the input and update the memory unit information and pass it to the next hidden layer unit, and finally output the position coordinate information at time t+1.

In the training of the VR head-mounted device trajectory prediction model, it is necessary to predict the longer future trajectory of the VR head-mounted device based on the previous period of trajectory data. This requires the LSTM network structure to be continuously cycled, and the t+1 time The output is then used as the input at time t+2. In this loop, the prediction of long-term VR head-mounted device trajectory information can be completed. Experiments were conducted using the position coordinates of the VR head-mounted device with various step sizes as network input. During the network training process, the network evaluates and optimizes the output predicted value based on the known trajectory information of the VR head-mounted device, calculates the average three-dimensional space distance between the predicted value and the actual value as the training target, and selects the predicted value The mean square error with the actual value is used as the loss function, and the back propagation algorithm is used to update the weight parameters of the LSTM model based on the calculated loss function gradient. The training process is repeated until the preset training times are reached or until convergence.

In step S2 of this embodiment, the sensor used by the VR head-mounted device to detect human posture and gravity acceleration selects the InvenSense MPU6050 chip that integrates a 3-axis gyroscope and 3-axis acceleration. In order to better observe the acceleration changes, the acceleration in the three directions is analyzed separately, and the X, Y, Z three-phase coordinate axes are established; in order to facilitate the analysis between attitude angles, the MPU6050 chip is used to collect human body data, which has It has a three-directional coordinate system and can measure attitude angles, which are Pitch angle, Roll angle and Yaw angle. As shown in Figure 2, the yaw angle represents the offset angle from the Y-axis; the pitch angle represents the offset angle from the X-axis; and the roll angle represents the deflection angle from the Z-axis. The X-axis is perpendicular to the left and right sides of the human body, the Z-axis is perpendicular to the front and rear sides of the human body, and the Y-axis is perpendicular to the X-axis and Z-axis. By installing the MPU6050 chip on the VR head-mounted device and establishing this posture coordinate system. Acceleration is the acceleration AccX, AccY, AccZ along the X, Y, and Z axes.

The fall prediction model in step S3 includes an attitude angle distinguishing unit and an acceleration distinguishing unit. The fall prediction model executes the judgment process shown in Figure 3.

After the posture angle is input into the posture angle distinction unit, formula (1) can be used to distinguish the forward bending and backward falling of the human body from other postures, and obtain sqrtRP;

sqrtRP calculates RPP_delta according to formula (2), and compares it with the set RPP_delta threshold, which can effectively extract the sideways and backward bending and falling movements of the human body.

RPP_delta=fabs(SqrtRP-yaw) (2)

When the human body is standing normally, the RPP_delta angle does not exceed 30°. When bending or falling occurs, the RPP_delta angle exceeds 30°. Therefore, the RPP_delta threshold angle can be set to 30°. When the RPP_delta angle is less than 30°, it is considered to be the normal activity state of the human body; when the RPP_delta angle is greater than 30°, the acceleration data is input into the acceleration differentiation unit for the next step of discussion and judgment on the correlation of gravity acceleration. .

Since the bending and falling movements have the same changes in posture angle, all the normal bending and falling postures need to be distinguished from the intensity of their movements. The intensity of the movements needs to be judged by the acceleration of gravity.

When falling or leaning back, the yaw angle will decrease. When the yaw angle drops to the threshold value of 70°, the acceleration differentiation unit first calculates the Z-axis according to the acceleration in the Z-axis direction and the acceleration in the Y-axis direction according to formula (3) The combined acceleration with the Y axis: SqrtAccYZ. Then ACCYZZ_delta is calculated based on SqrtAccYZ, the acceleration in the Z-axis direction and formula (4), and then compared with the normal threshold and fall threshold to determine whether the human body has fallen;

AccY and AccZ are the accelerations in the Y and Z directions respectively; the Z-axis acceleration can reflect whether there is violent backward movement at this moment, whether it is weightless, or a violent impact with the ground and the seat; at the same time, when the human body falls backward, Y The downward acceleration in the axis direction will also change significantly.

AccYZZ_delta=fabs(SqrtAccYZ-AccZ) (4)

Observe the changes in ACCYZZ_delta angle. When leaning back and lying down, the ACCYZZ_delta amplitude is about 80. When a fall occurs, the ACCYZZ_delta amplitude is about 40. Therefore, the normal threshold of ACCYZZ_delta can be set to 80. If ACCYZZ_delta is greater than 80, it is judged as leaning back or lying back. When lying back, set the fall threshold of ACCYZZ_delta to 40. When ACCYZZ_delta is less than 40, it is judged as a fall. This further distinguishes normal actions such as leaning back from falling actions.

Example 2

After predicting the fall position of the human body, in this embodiment, the companion robot is configured to go to the position coordinates predicted in step S5 for protection.

The companion robot basically consists of the following components: chassis, wheels, battery, DC motor, DC motor driver module, Bluetooth communication module, core controller module, sensors, and memory foam pads. Memory foam has excellent energy absorption, can effectively absorb impact and energy, and reduce the impact of a fall; it can provide better shock absorption effect, gradually adapt to the shape of the head, and slow down the spread of impact; it can be adjusted according to the shape and weight of the head Provides personalized support and therefore better support in the event of a fall.

The companion robot uses the path planning algorithm (A* algorithm) to calculate the robot's path from the current location to the target location.

The A* algorithm uses heuristic functions to search for paths. The A* algorithm is suitable for path planning with known global environmental information. During the operation of the algorithm, two lists will be generated, Openlist and Closelist. Openlist contains nodes that have been visited but have not been selected as the next node. Closelist contains nodes that have not been visited or have not been expanded after being visited.

The A* algorithm performs path planning through the cost estimation function, extending from the initial position to nearby child nodes. The cost estimation function will select the node with the smallest cost value as the next parent node based on the distance cost of the child node, and repeat this process until the path is completed. Planning, generating the final path.

The general expression of the heuristic function is:

f(n)=g(n)+h(n)

In the formula, f(n) represents the cost estimate from the starting grid point to the target point via any node n, g(n) is the actual distance evaluation from the starting point to grid n, h(n) represents the longest distance from grid n to the end point. Distance assessment of optimal routes.

The path planning process is shown in Figure 4:

(1) Set up Openlist and Closelist. Openlist is used to store nodes that have been searched, and Closelist is used to store unexplored nodes.

(2) Set Nx as the initial grid point and Ny as the specified target point. Place initial grid points in Openlist. (3) Visit nodes near the initial grid point, place non-obstacle points in the Openlist, use the heuristic function to calculate their estimated distance and place the initial grid point in the Closelist.

(4) Place the next expansion point N in the Closelist according to the optimal selection strategy. If N is the specified target raster, the path planning ends and the algorithm runs successfully. If not, proceed to the following steps.

(5) Continue to execute the strategy of visiting its surrounding nodes for grid N. For surrounding sub-nodes that are not in Closelist, add them to Openlist for cost calculation.

(6) Return to step (4) until expanded to the target point.

(7) The algorithm runs successfully, the map and extended nodes are saved, and the backtracking connection is the final path.

The position of the companion robot in the experience space can be obtained by installing a radar in the corner of the room and establishing a navigation coordinate system.

Claims (7)

1. A fall position identification method, characterized by comprising the steps of:

s1, training an LSTM neural network;

and (3) data collection: collecting a time sequence of position coordinate data of the vr head-mounted device, wherein the position coordinate data of each time step in the time sequence comprises coordinates (xi, yi, zi) of the vr head-mounted device in a navigation coordinate system, i is a number of the time step, x is a coordinate in the northbound direction, y is a coordinate in the northbound direction, and z is a coordinate perpendicular to the horizontal plane;

data preparation: respectively carrying out standardization processing on the numerical values corresponding to the position coordinate data of each time step in the time sequence in the three coordinate axes of the navigation coordinate system, so that the mean value of the numerical values corresponding to the position coordinate data of each time step in the time sequence after the standardization processing in the three coordinate axes of the navigation coordinate system is 0, and the variance is 1;

dividing the data set: dividing the time sequence after the normalization processing into an input sequence and a target sequence;

LSTM neural network training: initializing LSTM model parameters, performing forward propagation on an input sequence through the LSTM model to obtain a predicted sequence, calculating a loss function according to the predicted sequence and a target sequence, updating weight parameters of the LSTM model according to the calculated loss function gradient through a reverse propagation algorithm, and repeating the training process until the preset training times are reached or until convergence is achieved;

s2, taking the ground as a reference, establishing a navigation coordinate system in a VR experience space, and collecting real-time data of the VR head-mounted device when a human body wears the VR head-mounted device, wherein the real-time data comprise attitude angles, acceleration and position coordinate data in the navigation coordinate system;

s3, inputting the attitude angle and the acceleration into a falling pre-judging model, and extracting real-time data triggering falling conditions;

s4, extracting position coordinate data corresponding to time steps in the sliding window range from the real-time data extracted in the step S3 by adopting a sliding window method to form a sample time sequence, wherein the time steps in the sliding window range comprise the time steps of the real-time data extracted in the step S3;

s5, carrying out standardization processing on the S4 sample time sequence, inputting the standardized LSTM neural network, and carrying out inverse standardization processing on the output result to obtain a prediction result of the position coordinate data of the VR head-mounted device.

2. The fall position identification method according to claim 1, wherein the step of normalizing comprises:

respectively solving the mean value and the variance of the numerical values corresponding to the position coordinate data of all the time steps in three coordinate axes of a navigation coordinate system;

converting the numerical value corresponding to the position coordinate data of each time step in the time sequence in three coordinate axes of the navigation coordinate system into a standardized numerical value by using the following formula:

normalized value = (original value-mean)/variance;

the original value is a numerical value corresponding to one coordinate axis selected by the position coordinate data of each time step in the navigation coordinate system, the average value is a mean value of numerical values corresponding to the selected one coordinate axis of the position coordinate data of all time steps in the navigation coordinate system, and the variance is a variance of numerical values corresponding to the selected one coordinate axis of the position coordinate data of all time steps in the navigation coordinate system.

3. A fall position identification method as claimed in claim 1, characterized in that, when dividing the data set, a sliding window method is used, in which the time sequence after normalization within one time window is taken as the input sequence, and the time sequence of the next time step of the input sequence is taken as the target sequence.

4. A fall position identification method as claimed in claim 1, characterized in that, when dividing the data set, a sliding window method is used to divide the time sequence after normalization processing within one time window into time sequences of three successive time steps, the time sequence of the first two time steps being the input sequence and the time sequence of the last time step being the target sequence.

5. The fall position recognition method according to claim 1, wherein the posture angle in step S2 includes a Pitch angle, a Roll angle, a Yaw angle, the Yaw angle representing an offset angle from the Y axis; the pitch angle represents the offset angle from the X-axis; the roll angle represents the deflection angle with the Z axis, the X axis is the axis perpendicular to the left and right sides of the human body, the Z axis is the axis perpendicular to the front and back sides of the human body, the Y axis is perpendicular to the X and Z axes, and the acceleration is the acceleration AccX, accY, accZ along the X, Y, Z axis.

6. The fall position recognition method according to claim 1, wherein the fall pre-determination model in step S3 includes an attitude angle distinguishing unit and an acceleration distinguishing unit, and the fall pre-determination model performs the steps of determining whether a fall condition is triggered:

after the attitude angle is input into the attitude angle distinguishing unit, the forward bending and backward falling of the human body can be distinguished from other attitudes by utilizing the formula (1) to obtain the sqrtRP;

the sqrtRP calculates RPP_delta according to the formula (2), and compares the RPP_delta with a set RPP_delta threshold value, so that the bending and falling actions of the human body in the lateral and backward directions can be effectively extracted.

RPP_delta＝fabs(SqrtRP-yaw) (2)

Inputting the acceleration into the acceleration distinguishing unit when the RPP_delta angle is larger than 30 degrees;

when the yaw angle drops to not more than 70 °, the acceleration distinguishing unit first calculates the resultant acceleration of the Z axis and the Y axis according to formula (3) based on the acceleration of the Z axis direction and the acceleration of the Y axis direction: sqrtAccYZ, calculating ACCYZZ_delta according to the SqrtAccYZ, the acceleration in the Z axis direction and the formula (4), and comparing the ACCYZZ_delta with a normal threshold value and a falling threshold value to judge whether the falling action of the human body occurs;

AccYZZ_delta＝fabs(SqrtAccYZ-AccZ) (4)

observing the change condition of ACCYZZ_delta angle, and judging that the patient falls when ACCYZZ_delta is smaller than 40, and triggering a falling condition.

7. A companion robot, characterized in that the companion robot goes to the position coordinates predicted in step S5 of claim 1 to be protected, and the companion robot carries a memory foam cushion for absorbing impact and energy and reducing the impact force of the head when falling down.