Human-machine interaction system of quad-rotor unmanned aerial vehicle based on fine gesture recognition

Abstract

The application discloses a four-rotor unmanned aerial vehicle human-computer interaction system based on fine gesture recognition, belongs to the technical field of unmanned aerial vehicle control, and solves the problems of low accuracy of an existing fine gesture recognition algorithm and insufficient robustness of an existing PID control algorithm in a dynamic scene; the technical scheme of the application adopts a feature fusion network based on a hollow residual error to perform gesture recognition on data transmitted by an IMU sensor glove, feeds shallow features into a hollow residual network to extract shallow features and deep features and perform feature fusion, so that the extracted features are more reliable and fine, contain not only shallow detail information and deep high semantic information, and noise information is reduced, and the fine gesture recognition accuracy is improved; and the parameters of a cascade PID controller are dynamically adjusted by using reinforcement learning, so that the robustness of the PID control is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of unmanned aerial vehicle control and relates to a four-rotor unmanned aerial vehicle human-computer interaction system based on fine gesture recognition. BACKGROUND

[0002] In the past decade, the use of unmanned aerial vehicles (UAVs) has increased dramatically, with applications ranging from package delivery services, military uses, and disaster management to photography and entertainment.

[0003] Gesture is a way of human-computer interaction, which is widely studied due to its convenience and user-friendliness. However, the concealment of gesture interaction is still a major application requirement for commercial and military use, and further research is needed. IMU (Inertial Measurement Unit) sensors are a medium for gesture interaction, and compared with visual methods, IMU sensors are not affected by equipment and light, and can be used for gesture interaction in concealed environments. However, the common gesture interaction method in complex concealed environments is easily discovered, and the fine gesture recognition accuracy for achieving concealed interaction is low.

[0004] In the design of a four-rotor unmanned aerial vehicle, the design of the four-rotor unmanned aerial vehicle controller is an important part of the four-rotor unmanned aerial vehicle to achieve trajectory tracking. The PID controller is simple in structure, easy to implement, and suitable for most control objects, so it is widely used. In a concealed environment, the existing PID control method lacks robustness and cannot achieve dynamic interaction with the unknown dynamic environment. Different PID parameters result in different effects of the control system, and changes in the external environment result in poor effects of the traditional control parameter tuning method. It is difficult for a set of PID parameters to cover the entire flight envelope, and it is necessary to improve the environmental adaptability and interaction ability of the PID parameter tuner. Therefore, control parameter tuning is an important problem of the PID controller. SUMMARY

[0005] The technical problem to be solved by the application is how to design a four-rotor unmanned aerial vehicle human-computer interaction system based on fine gesture recognition to solve the problems of low accuracy of existing fine gesture recognition algorithms and insufficient robustness of existing PID control algorithms in dynamic scenes.

[0006] The application solves the above technical problems by the following technical solutions:

[0007] The application discloses a control method of a four-rotor unmanned aerial vehicle system based on fine gesture recognition, and the four-rotor unmanned aerial vehicle system comprises a user terminal and a PC terminal; a user of the user terminal wears an IMU sensor glove in a concealed environment, observes the environment of the unmanned aerial vehicle in real time, and controls the flight of the unmanned aerial vehicle by adopting corresponding fine gestures; the IMU sensor glove is used for capturing the fine gestures of the user and sending the fine gestures to an environment interaction module; the PC terminal comprises the environment interaction module and an adaptive cascade PID control module.

[0008] The control method is as follows:

[0009] The environment interaction module adopts a feature fusion network based on a hollow residual error to perform gesture recognition on the data sent by the IMU sensor glove, and sends the recognized gesture instructions to the adaptive cascade PID control module.

[0010] The adaptive cascade PID control module sets the position coordinates to be tracked according to the received gesture instructions, performs trajectory tracking by using a position-based cascade PID control system, and converts the parameters of the cascade PID control system into an action selection function by using a continuous action reinforcement learning machine, so as to dynamically select the action and select the parameters of the cascade PID control system, so as to realize the setting of the PID parameters.

[0011] Further, the method for the environment interaction module to perform gesture recognition on the data sent by the IMU sensor glove based on the feature fusion network of the hollow residual error is as follows:

[0012] (1) a CNN convolution kernel is used to extract the features between signals, and a BN layer is used to normalize the features to obtain shallow features;

[0013] (2) a "hole" in the hollow convolution is used to obtain the fusion features between sensors, and a residual network is used to avoid network overfitting or gradient descent, and the specific formula is as follows:

[0014]

[0015] wherein the hollow rate r is used to expand the receptive field, V dcnn is an output feature;

[0016] After two layers of hollow convolution (r=1, r=2), a residual error D(X) is obtained, and the residual error obtained by the hollow convolution is connected with the feature X by jumping, and the formula is as follows:

[0017] R(X) = (X) +

[0018] wherein the input is pre-activated and pre-normalized by a linear unit function and BatchNorm.

[0019] (3) The shallow features are spliced and fused with the deep features, and the fused features F(X) are as follows:

[0020] F(X)=Cat(D low (),D high ())

[0021] Wherein, D low () represents the shallow feature, D high () represents the deep feature, and CAT() is a feature splicing function;

[0022] (4) The time sequence relationship between the fused features is learned through the GRU neural network, the information is selected and discarded by using the gate unit, the features learned by the GRU neural network are subjected to a full connection layer and a Softmax function, and a fine gesture recognition classification result is obtained, so that the process of fine gesture recognition is completed.

[0023] Further, the position-based cascade PID control system comprises a position cascade PID controller, a posture calculation module, a posture angle cascade PID controller and a reinforcement learning parameter setting module.

[0024] (1) The output of the position cascade PID controller comprises two parts: a horizontal position channel and a height channel; the horizontal position channel comprises an outer ring horizontal channel and an inner ring horizontal channel; the speed expected value of the outer ring horizontal channel is obtained through a proportional controller, and the inner ring horizontal channel uses a proportional, integral and differential controller to obtain, and the specific process is as follows:

[0025] V hd =K ph e h

[0026]

[0027] Wherein, e h =P hd -P h , P hd =[x d y d ] T , P h =[x y] T ; K ph =diag(k px ,k py ), e vh =[e vx e vy ] T , Kpvh = diag(k pvx ,k pvy ); K ivh = diag(k ivx ,k ivy ), K dvh = diag(k dvx ,k dvy );

[0028] The output of the height channel is shown as follows:

[0029] v zd = K pz e z

[0030]

[0031] wherein e Z = Z d - Z r , e vz = v zd - v z ;

[0032] The 12 PID parameters of the position cascade PID controller that need to be tuned are respectively k px , k py , k pvx , k pvy , k ivx , k ivy , k dvx , k dvy , K pz , K pvz , K ivz , K dvz ;

[0033] (2) The controller outputs u x , u y , and u z are obtained through the position cascade PID controller, and given ψ d , the given attitude angle and the height channel output are obtained by solving the attitude of the position dynamics model of the quadrotor unmanned aerial vehicle, and the calculation formula of the attitude solving module is as follows:

[0034]

[0035] (3) For the given attitude angle d in the above formula, through the attitude angle cascade PID controller, through the attitude angle outer loop and the angular velocity inner loop, the following are obtained respectively:

[0036] ω d = KpΩ e Ω

[0037]

[0038] where e Ω = Ω d - Ω, Ω d = [φ d θ d ψ d ] T are the desired attitude angles; e ω = ω d - ω, U d = [u2 u3 u4] T ; K pΩ , K pω , K iω , K dω are the parameters needed to be tuned for the attitude angle cascade PID controller, K pΩ = diag(k pφ , k pθ , k pψ ), K pω = diag(k pωφ , k pωθ , k pωψ ), K iω = diag(k iωφ , k iωθ , k iωψ ), K dω = diag(k dωφ , k dωθ , k dωψ ) are 12 PID parameters;

[0039] (4) The above 24 PID parameters are tuned by using a reinforcement learning parameter tuning module.

[0040] Further, the method for using a continuous action reinforcement learning machine to convert cascade PID control system parameters into an action selection function, taking minimizing a loss function as an objective, and dynamically selecting actions to select cascade PID control system parameters is as follows:

[0041] (1) The position error (e x , e y , e z ) and the attitude angle error (e θ , e φ , e ψ ) are taken as the input of the reinforcement learning parameter tuning module, and the action selection function is initialized as follows:

[0042]

[0043] (2) The loss function is calculated as follows:

[0044] J j =∫0 t τ[e j (τ)] 2 dτ

[0045] Wherein, e j ,j=0-6, respectively, the input e x ,e y ,e z ,e θ ,e φ ,e ψ ;

[0046] (3) The score is calculated as follows:

[0047]

[0048] (4) A random number in the range of (0, 1) is generated Solve k i that makes the equation true, and make action selection; wherein, k i 24 series PID parameters need to be adjusted in embodiment four, i=1~24;

[0049] (5) The parameters of the series PID controller are updated;

[0050] (6) The action selection function is updated as follows:

[0051]

[0052] (7) The number of iterations t is updated, when t>max_step, the PID parameter setting is ended, and the optimal parameters are obtained; otherwise, return to step (2).

[0053] Further, the fine gesture includes six gestures, which are respectively left fly, right fly, front fly, back fly, up fly and down fly.

[0054] Further, the IMU sensor glove sends data to the environment interaction module through a USB serial port or Bluetooth.

[0055] Further, the IMU sensor glove adopts a 6-axis IMU sensor.

[0056] The advantages of the present application are:

[0057] (1) The technical scheme of the present application comprises a control system of a four-rotor unmanned aerial vehicle system based on fine gesture recognition, which uses an inertial measurement unit as a medium for gesture interaction to realize human-computer interaction in a concealed environment. The gesture recognition algorithm based on hollow convolution solves the problems of low accuracy of existing fine gesture recognition algorithms and insufficient robustness of existing PID control algorithms in dynamic scenes.

[0058] (2) The technical scheme of the present application designs a flight control method of a tilt-rotor unmanned aerial vehicle based on reinforcement learning, which realizes dynamic interaction with an unknown dynamic environment. A position command-oriented cascade PID control system is designed, and a continuous action reinforcement learning machine is used to convert the cascade PID control system parameters into an action selection function. The goal is to minimize the loss function, and the action selection is performed dynamically to realize real-time online self-tuning of the cascade PID control system parameters. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The figure is a block diagram of the four-rotor unmanned aerial vehicle human-computer interaction system based on fine gesture recognition of the embodiment of the present application.

[0060] Figure 2 The figure is a flowchart of the control method of the four-rotor unmanned aerial vehicle human-computer interaction system based on fine gesture recognition of the embodiment of the present application.

[0061] Figure 3 The figure is a block diagram of the gesture recognition algorithm of the feature fusion network based on hollow residual of the embodiment of the present application.

[0062] Figure 4 The figure is a block diagram of the position-based cascade PID unmanned aerial vehicle control system of the embodiment of the present application.

[0063] Figure 5 The figure is a data flow diagram of the double-loop PID control parameter tuning based on reinforcement learning of the embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described below in conjunction with the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0065] The technical scheme of the present application will be further described below in conjunction with the drawings of the specification and specific embodiments:

[0066] Embodiment one

[0067] As Figure 1As shown in the figure, it is a four-rotor unmanned aerial vehicle human-computer interaction system block diagram based on fine gesture recognition, the four-rotor unmanned aerial vehicle system based on fine gesture recognition comprises: a user end and a PC end. The user end comprises: an IMU sensor glove and six specified unmanned aerial vehicle operation gestures (left fly, right fly, forward fly, backward fly, upward fly, downward fly); the IMU sensor glove comprises a total of six six-axis IMU sensors on the fingers and the back of the hand, which are used to capture the fine gestures of the user (fine gestures refer to the small, fine and difficult-to-distinguish gestures made by the hands and fingers according to specific gesture patterns, which are not specified here); the user as the main part of the unmanned aerial vehicle system is a key link in human-computer interaction, the user wears the IMU sensor glove, observes the environment where the unmanned aerial vehicle is located in real time, and controls the flight of the unmanned aerial vehicle by adopting the corresponding unmanned aerial vehicle operation gesture. The PC end comprises: a gesture recognition algorithm module and a four-rotor unmanned aerial vehicle control algorithm module, the gesture recognition algorithm is used to recognize the obtained user gesture data, obtain the corresponding gesture instruction, and call the four-rotor unmanned aerial vehicle control algorithm module to perform corresponding trajectory tracking.

[0068] As shown in the figure, it is a four-rotor unmanned aerial vehicle system based on fine gesture recognition, and the specific steps are as follows: Figure 2

[0069] Step S2-1: In a concealed environment, the user uses the IMU sensor glove worn on the hand to control the four-rotor unmanned aerial vehicle according to the corresponding four-rotor unmanned aerial vehicle gesture according to the scene; wherein the IMU sensor glove sends IMU data to the PC end through a USB serial port or Bluetooth for communication between data.

[0070] Step S2-2: The PC end receives the sensor signal, calls the gesture recognition algorithm module, obtains the corresponding gesture instruction, and sends it to the four-rotor unmanned aerial vehicle; specifically, after the IMU sensor data is filtered and normalized, it is input to the feature fusion network based on the hollow residual to recognize the gesture of the IMU data, obtain the corresponding gesture instruction, and send it to the unmanned aerial vehicle.

[0071] Step S2-3: The unmanned aerial vehicle receives the gesture instruction, sets the position coordinates to be tracked, and uses the position-based cascade PID controller to perform trajectory tracking; wherein the PC end obtains the given position (x d ,y d ,z d ) through the gesture recognition algorithm module, which is the given position to be tracked by the four-rotor unmanned aerial vehicle; in order to more flexibly control the flight of the four-rotor unmanned aerial vehicle, the given position of the four-rotor unmanned aerial vehicle is set as the current position P = (x, y, z) of the four-rotor unmanned aerial vehicle plus the position offset P e ​The position offset of the unmanned aerial vehicle is controlled by gesture instructions in the form of P = (dx, dy, dz). The position offset increments of the unmanned aerial vehicle in each direction are (dx, dy, dz), which are given before the unmanned aerial vehicle flies, wherein (x, y, z) is the real position of the unmanned aerial vehicle, which is obtained by the GPS in the unmanned aerial vehicle. Specifically, when the user sends a left flying instruction, P e = (-dx, 0, 0); when the user sends a right flying instruction, P e = (dx, 0, 0); when the user sends a front flying instruction, P e = (0, dy, 0); when the user sends a rear flying instruction, P e = (0, -dy, 0); when the user sends an up flying instruction, P e = (0, 0, dz); when the user sends a rear flying instruction, P e = (0, 0, -dz); in the real scene of the unmanned aerial vehicle, z > 0, z-dz > 0.

[0072] Step S2-4: In order to cope with the PID parameter setting in the complex scene, a reinforcement learning algorithm is used to interact with the environment, obtain the corresponding score to select the action, and realize the PID parameter setting.

[0073] As shown in Figure 3 The application proposes a gesture recognition algorithm framework based on feature fusion of a hollow residual for fine gesture classification. In a hidden scene, the interactive action of the gesture and the unmanned aerial vehicle needs to be as small as possible, which causes a certain influence on the accuracy of gesture recognition. The traditional convolutional neural network mostly adopts the method of adding a pooling layer after a convolution layer, expands the receptive field, speeds up convergence, and extracts global information; but the addition of the pooling layer will lose the correlation in space and reduce the feature resolution. For fine gestures, high-resolution features are more helpful for fine gesture classification.

[0074] For the fine gesture data I of the application, the IMU sensor obtains signals on M = 36 axes in space and N frames of hand motion information in time sequence. The given gesture keystroke data is first subjected to a CNN convolution layer to obtain unique features X(x, y) between different signals (acceleration, angular velocity and attitude angle signals). For the 6-axis IMU sensor used in the application, a 3*3 convolution kernel is used to extract the features between signals, and the step is 3; and a BN layer is used to normalize the features to obtain shallow features.

[0075] Secondly, the hollow convolution is used to obtain the fusion features between the sensors by using the "hole" in the hollow convolution, and the residual network is used to avoid overfitting or gradient descent of the network, and the specific formula is as follows:

[0076]

[0077] wherein the hole rate is r to enlarge the receptive field, V dcnn is the output feature;

[0078] After two layers of hole convolution (r=1, r=2), residual D(X) is obtained, and the residual obtained by the hole convolution is connected with the feature X by jumping, as shown in the following formula:

[0079] R(X) = (X) +

[0080] wherein the input is pre-activated and pre-normalized by a linear unit function (Rectified Linear Unit, ReLu) and BatchNorm.

[0081] Again, the shallow feature has the characteristics of high resolution and contains more detailed information, but the semantic of the shallow feature is low and the noise is large. The deep feature contains high semantic information, but the resolution is low. The shallow feature and the deep feature are fused, so that the shallow feature and the deep feature are complementary, not only containing more detailed information, but also having high semantic information. The shallow feature D low () and the deep feature D ig h() are spliced, and the fused feature F(X) is as shown in the following formula:

[0082] F(X) = Cat(D low (), D high ())

[0083] wherein D low () represents the shallow feature, D high () represents the deep feature, and CAT() is a feature splicing function.

[0084] Finally, the GRU neural network learns the time sequence relationship between the fused features, and uses the gate unit to select and discard information, so as to solve the long period dependence problem. The feature learned by the GRU neural network is subjected to a full connection layer and a Softmax function, so as to obtain a fine gesture recognition classification result.

[0085] As shown in Figure 4 , the embodiment of the present application adopts a position-based cascade PID control system to track the trajectory of the quadrotor unmanned aerial vehicle; the position-based cascade PID control system comprises a position cascade PID controller, an attitude calculation module, an attitude angle cascade PID controller, a reinforcement learning parameter setting module and a quadrotor unmanned aerial vehicle.

[0086] Firstly, the position cascade PID controller includes two parts: horizontal position channel and height channel; the horizontal position channel includes outer loop horizontal channel and inner loop horizontal channel; the speed expectation value of the outer loop horizontal channel is obtained by a proportional controller, and the inner loop horizontal channel uses proportional, integral and differential controllers, which are as follows:

[0087] V hd =K ph e h

[0088]

[0089] wherein e h =P hd -P h , P hd =[x d y d ] T , P h =[x y] T ; K ph =diag(k px ,k py ), e vh =[e vx e vy ] T , K pvh =diag(k pvx ,k pvy ); K ivh =diag(k ivx ,k ivy ), K dvh =diag(k dvx ,k dvy ).

[0090] The height channel is similar to the horizontal channel, which is as follows:

[0091] v zd =K pz e z

[0092]

[0093] wherein e Z =Z d -Z r , e vz =v zd -v z .

[0094] The 12 PID parameters of the position cascade PID controller which need to be tuned are as follows: kpx , k py , k pvx , k pvy , k ivx , k ivy , k dvx , k dvy , K pz , K pvz , K ivz , K dvz .

[0095] Secondly, the controller output u x , u y , u z is obtained by the position cascade PID controller, and given ψ d , the given attitude angle and the height channel output can be obtained by solving the attitude of the position dynamics model of the quadrotor unmanned aerial vehicle, and the attitude solving is as follows:

[0096]

[0097] Next, for the given attitude angle of the above formula d , through the attitude angle cascade PID controller, through the attitude angle outer loop and the angular velocity inner loop, the following can be obtained:

[0098] ω d = K pΩ e Ω

[0099]

[0100] Wherein, e Ω = Ω d - Ω, Ω d = [φ d θ d ψ d ] T is the desired attitude angle; e ω = ω d - ω, U d = [u2 u3 u4] T ; K pΩ , K pω , K iω , K dω , are the parameters required to be adjusted by the attitude angle cascade PID controller, K pΩ = diag(k pφ , k pθ , k pψ ), K pω = diag(k pωφ , k pωθ , kpωψ ),K iω =diag(k iωφ ,k iωθ ,k iωψ ),K dω =diag(k dωφ ,k dωθ ,k dωψ ), a total of 12 PID parameters.

[0101] Finally, in order to solve the problem of low robustness of PID parameters in complex environment, a reinforcement learning parameter setting module is used to set the above-mentioned 24 PID parameters.

[0102] The method for setting the above-mentioned 24 PID parameters by using the reinforcement learning parameter setting module is as follows: using a continuous action reinforcement learning automata (CARLA, Continuous Action Reinforcement Learning Automata), the cascade PID parameters are converted into an action selection function, the loss function is minimized as the goal, and the action selection is dynamically performed, so as to select the cascade PID parameters; specifically, the input position error (e x ,e y ,e z ) and the attitude angle error (e θ ,e φ ,e ψ ) are taken as the input of the reinforcement learning, and the specific steps are as follows:

[0103] Step S5-1: initialize the action selection function (t=1);

[0104]

[0105] Step S5-2: calculate the loss function

[0106] J j =∫0 t τ[e j (τ)] 2 dτ

[0107] Wherein, e j ,j=0-6, are the inputs of the reinforcement learning e x ,e y ,e z ,e θ ,e φ ,e ψ ;

[0108] Step S5-3: calculate the system score

[0109]

[0110] Step S5-4: generating a random number in the range of (0, 1) Solving k that makes the equation true i , action selection is performed; wherein, k i are 24 series PID parameters that need to be tuned in the fourth embodiment, i = 1-24;

[0111] Step S5-5: updating the parameters of the series PID controller of the quadcopter unmanned aerial vehicle;

[0112] Step S5-6: updating the action selection function

[0113]

[0114] Step S5-7: updating the iteration number (t = t + 1), when t > max_step, the PID parameter tuning is completed, and the optimal parameters are obtained; otherwise, return to step S5-2.

[0115] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A four-rotor unmanned aerial vehicle human-computer interaction system based on fine gesture recognition, characterized in that, The system comprises a user terminal and a PC terminal; a user of the user terminal wears an IMU sensor glove in a concealed environment, observes the environment where the unmanned aerial vehicle is located in real time and adopts corresponding fine gestures to control the flight of the unmanned aerial vehicle, the IMU sensor glove is used to capture the fine gestures of the user and send them to an environment interaction module; the PC terminal comprises the environment interaction module and an adaptive cascade PID control module; The control method of the four-rotor unmanned aerial vehicle human-computer interaction system is as follows: The environment interaction module adopts a feature fusion network based on a hollow residual error to recognize the data sent by the IMU sensor glove and sends the recognized gesture instructions to the adaptive cascade PID control module; The adaptive cascade PID control module sets the position coordinates to be tracked according to the received gesture instructions, uses a position-based cascade PID control system to perform trajectory tracking, and uses a continuous action reinforcement learning machine to convert the cascade PID control system parameters into an action selection function, with the aim of minimizing a loss function, dynamically selects actions, and thus selects the cascade PID control system parameters, so as to realize the setting of the PID parameters; The position-based cascade PID control system comprises a position cascade PID controller, an attitude calculation module, an attitude angle cascade PID controller and a reinforcement learning parameter setting module; the method for using the position-based cascade PID control system to perform trajectory tracking is as follows: The output of the position cascade PID controller comprises two parts: a horizontal position channel and a height channel; the horizontal position channel comprises an outer ring horizontal channel and an inner ring horizontal channel; the speed expectation value of the outer ring horizontal channel is obtained through a proportional controller, and the inner ring horizontal channel uses a proportional, integral and differential controller to obtain, and the specific process is as follows: wherein ; ; , ; The output of the height channel is as follows: wherein ; The 12 cascade PID parameters of the position cascade PID controller are obtained, which are respectively: , , , , , , , , , , , ; (2) Obtain the controller output through the position cascade PID controller , , and given , the given attitude angle and height channel output are obtained by solving the attitude of the quad-rotor unmanned aerial vehicle dynamics model, and the calculation formula of the attitude solving module is described as follows: (3) For the given attitude angle Through the attitude angle cascade PID controller, through the attitude angle outer loop and angular velocity inner loop, respectively wherein, , is the desired attitude angle; , ; is the parameter to be tuned for the attitude angle cascade PID controller, , 12 cascade PID parameters in total; (4) The 24 series PID parameters are set by using a reinforcement learning parameter setting module, and the 24 series PID parameters are as follows: , , , , , , , , , , , , , , , , , , , , , , , .

2. The fine gesture recognition based quadcopter drone human-machine interaction system of claim 1, wherein, The method for the environment interaction module to recognize the data sent by the IMU sensor glove based on a hollow residual error is as follows: (1) a CNN convolution kernel is used to extract the features between signals, and a BN layer is used to normalize the features to obtain shallow features; (2) a "hole" in the hollow convolution is used to obtain the fusion features between sensors, and a residual network is used to avoid network overfitting or gradient descent, and the specific formula is as follows: Wherein, the void ratio is to enlarge the receptive field, is an output feature; After two layers of void convolution, i.e. Obtain the residual The residuals and features obtained from dilated convolution are combined. Perform a jump connection as shown in the following equation: where the input is pre-activated and pre-normalized through a linear unit function and BatchNorm; (3) The shallow features are spliced and fused with the deep features, and the fused features as shown in the following formula: wherein, represents a shallow feature, represents a deep feature, Cat () is a feature concatenation function; (4) the time sequence relationship between the fusion features is learned through a GRU neural network, the information is selected and discarded through a gate unit, the features learned by the GRU neural network are processed through a full connection layer and a Softmax function, the classification results of the fine gestures are obtained, and thus the process of fine gesture recognition is completed.

3. The fine gesture recognition based quadcopter human-machine interaction system according to claim 2, wherein, The method for using the continuous action reinforcement learning machine to convert the cascade PID control system parameters into an action selection function, with the aim of minimizing a loss function, dynamically selecting actions, and thus selecting the cascade PID control system parameters is as follows: (1) The position error and the attitude angle error are taken as the input of the reinforcement learning parameter setting module, and the action selection function is initialized as follows: (2) the loss function is calculated as follows: wherein, 0-5, respectively, are inputs for reinforcement learning ; (3) The score is calculated as follows: (4) generating a random number in the range of (0, 1) making the equation true , action selection is performed; wherein, are 24 series PID parameters that need to be tuned, ; (5) The parameters of the cascade PID controller are updated; (6) The action selection function is updated as follows: (7) The iteration number t is updated, when t>max_step, the PID parameter setting is ended, and the optimal parameters are obtained; otherwise, return to step (2).

4. The fine gesture recognition based quadcopter human-machine interaction system of claim 1, wherein, The fine gesture includes six gestures, which are respectively left fly, right fly, front fly, back fly, up fly and down fly.

5. The fine gesture recognition based quadcopter human-machine interaction system of claim 1, wherein, The IMU sensor glove sends data to the environment interaction module through a USB serial port or Bluetooth.

6. The fine gesture recognition based quadcopter drone human-machine interaction system of claim 1, wherein, The IMU sensor glove adopts a 6-axis IMU sensor.

Images