A vision-based method and system for driverless vehicle following, and the driverless vehicle itself.

Abstract

This invention relates to the field of autonomous driving technology, and more specifically, to a vision-based method, system, and autonomous vehicle for following people. This invention takes into account the lack of global path guidance during the person-following process and calculates the front wheel steering angle δ based on positional relationships. f (t), and with the actual distance L d (t) and expected distance L r The error e(t) is input into the PID controller, which outputs the autonomous vehicle's acceleration / deceleration signal acc(t). This signal controls the vehicle's acceleration and deceleration, thus achieving lateral and longitudinal control. Furthermore, to improve the autonomous vehicle's personnel following performance, this invention also uses the DDPG model to adjust the L... d The parameters of (t) and PID are adjusted, and the policy network parameters are updated by utilizing the self-iterative training of the autonomous vehicle during the process of following the personnel, thereby effectively improving the L... d The parameters of (t) and PID are adjusted to avoid tedious manual debugging.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and more specifically, to a vision-based method, system, and autonomous vehicle for following people. Background Technology

[0002] The research on autonomous driving control mainly consists of lateral control and longitudinal control. Lateral control primarily achieves target path tracking by controlling the steering angle of the front wheels; longitudinal control mainly improves tracking performance and driving safety by controlling vehicle speed.

[0003] Existing lateral control methods designed based on tracking error require calculations based on the global reference path and the current vehicle state. However, when an autonomous vehicle follows a person, the lack of a global path makes it difficult to obtain the tracking error, rendering existing lateral control methods unsuitable. Summary of the Invention

[0004] Therefore, it is necessary to address the problem that existing lateral control methods are not suitable for controlling unmanned vehicles to follow people, and to provide a vision-based method, system, and unmanned vehicle for people following.

[0005] This invention is achieved using the following technical solution:

[0006] In a first aspect, the present invention discloses a vision-based human following method for unmanned vehicles, which is used for lateral and longitudinal control of unmanned vehicles.

[0007] The vision-based human tracking method for autonomous vehicles includes the following steps:

[0008] Obtain the actual distance L between the unmanned vehicle and the target person at the current time t. d (t), the angle α(t) between the direction of the line connecting the rear axle of the unmanned vehicle to the target person at the current time t and the direction of the unmanned vehicle's movement;

[0009] Obtain the preset value K of the proportional gain P Integral gain preset value K I Differential gain preset value K D ;

[0010] Construct a DDPG network model and analyze L d (t), K P K I K D Processing yields L d ′(t), K′ P (t), K′ I (t), K′ D (t); where L d ′(t) is Ld The correction value of (t), K′ P (t) is K P Correction value, K′ I (t) is K I Correction value, K′ D (t) is K D Correction value;

[0011] The method of lateral control is as follows:

[0012] Calculate the front wheel angle of the unmanned vehicle at the current time t. And according to δ f (t) performs the steering; where L represents the wheelbase of the driverless vehicle;

[0013] The method of longitudinal control is as follows:

[0014] Compare L d (t) and the preset expected distance L r The error at the current time t is obtained as e(t) = L. d (t)-L r ;

[0015] The unmanned vehicle's speed is controlled based on PID control according to e(t) to maintain a safe distance between the unmanned vehicle and the target personnel; the PID parameter uses K′ P (t), K′ I (t), K′ D (t).

[0016] The driverless vehicle implements a vision-based person-following method according to embodiments of this disclosure.

[0017] Secondly, the present invention discloses a vision-based human follow system for unmanned vehicles, which uses the vision-based human follow method for unmanned vehicles disclosed in the first aspect.

[0018] The vision-based human-following system for autonomous vehicles includes: a parameter acquisition module, a DDPG network module, a lateral control module, and a longitudinal control module.

[0019] The parameter acquisition module is used to obtain the actual distance measurement value L between the unmanned vehicle and the target person at the current time t. d (t), the angle α(t) between the line connecting the rear axle of the unmanned vehicle to the target person at the current time t and the forward direction of the unmanned vehicle; also used to obtain the preset value K of the proportional gain. P Integral gain preset value K I Differential gain preset value K D .

[0020] The DDPG network module is used to build DDPG network models and to handle L... d(t), K P K I K D Processing yields L d ′(t), K′ P (t), K′ I (t), K′ D (t). Wherein, L d ′(t) is L d The correction value of (t), K′ P (t) is K P Correction value, K′ I (t) is K I Correction value, K′ D (t) is K D The correction value.

[0021] The lateral control module is used to calculate the front wheel steering angle of the unmanned vehicle at the current time t. And according to δ f (t) performs a turn. Where L represents the wheelbase of the autonomous vehicle.

[0022] The longitudinal control module is used to compare L d (t) and the preset expected distance L r The error at the current time t is obtained as e(t) = L. d (t)-L r The system controls the vehicle speed based on PID control according to e(t), ensuring a safe distance between the vehicle and the target personnel. The PID parameters are K′. P (t), K′ I (t), K′ D (t).

[0023] The driverless vehicle uses a vision-based human-following system to implement a method or process according to embodiments of this disclosure.

[0024] Thirdly, the present invention discloses an unmanned vehicle that uses the vision-based human following method for unmanned vehicles disclosed in the first aspect.

[0025] Fourthly, the present invention discloses a readable storage medium. The readable storage medium stores computer program instructions, which, when read and executed by a processor, perform the steps of the vision-based human-following method for unmanned vehicles disclosed in the first aspect.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] This invention takes into account the lack of global path guidance for autonomous vehicles during the process of following personnel, and calculates the front wheel steering angle δ based on positional relationships. f (t), and with the actual distance Ld (t) and expected distance L r The error e(t) is input into the PID controller, which outputs the autonomous vehicle's acceleration / deceleration signal acc(t). This signal controls the vehicle's acceleration and deceleration, thus achieving lateral and longitudinal control. To improve the driver-following performance of the autonomous vehicle, this invention uses the DDPG model to optimize the L... d The parameters of (t) and PID are adjusted, and the policy network parameters are updated by utilizing the self-iterative training of the autonomous vehicle during the process of following the personnel, thereby effectively improving the L... d The parameters of (t) and PID are adjusted to avoid tedious manual debugging. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the vision-based human following method for unmanned vehicles in Embodiment 1 of the present invention;

[0029] Figure 2 This is a schematic diagram showing the positional relationship between the unmanned vehicle and the target personnel in Embodiment 1 of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] It should be noted that when a component is said to be "installed on" another component, it can be directly on the other component or it may be in a component that is centered on it. When a component is said to be "set on" another component, it can be directly set on the other component or it may also be in a component that is centered on it. When a component is said to be "fixed to" another component, it can be directly fixed to the other component or it may also be in a component that is centered on it.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0033] Example 1

[0034] Please see Figure 1This is a schematic diagram of a vision-based human following method for an unmanned vehicle disclosed in Embodiment 1. The vision-based human following method for an unmanned vehicle in Embodiment 1 is used for lateral and longitudinal control of the unmanned vehicle. In the context of the prior art, lateral control refers to the control of the front wheel steering angle, and longitudinal control refers to the control of the vehicle speed.

[0035] See Figure 2 At the current time t, the autonomous vehicle and the following person are in the same XY coordinate system. The target person is located on one side of the autonomous vehicle. The autonomous vehicle has front wheels and rear wheels, with a wheelbase of L.

[0036] Setting X r (t), Y r (t) represents the horizontal and vertical coordinates of the rear axle center of the unmanned vehicle at the current time t, and X is set as follows: p (t), Y p (t) represents the horizontal and vertical coordinates of the target person at the current time t. The distance L between the target person and the unmanned vehicle is... d More specifically, L d This indicates the distance from the center of the rear axle of the autonomous vehicle to the target person.

[0037] In this embodiment 1, L d Take L d (t) The corrected value L obtained after processing by the DDPG network model d ′(t). Where, L d (t) represents the actual distance measurement between the unmanned vehicle and the target person at the current time t.

[0038] Additionally, the heading angle of the unmanned vehicle at the current time t is φ(t). The angle between the direction of the line connecting the rear axle of the unmanned vehicle to the target person and the direction of travel of the unmanned vehicle at the current time t is α(t). There is a conversion relationship between α(t) and φ(t):

[0039]

[0040] like Figure 2 Based on the Law of Cosines, we can obtain: After conversion, the result is Where R is the instantaneous turning radius.

[0041] Furthermore, according to the vehicle's kinematic equations: Then there is Where, δ f (t) represents the front wheel steering angle of the unmanned vehicle at the current time t; L represents the wheelbase of the unmanned vehicle.

[0042] Regarding vehicle speed, the PID controller adjusts the speed of the autonomous vehicle based on e(t) to maintain a safe distance between the vehicle and the target personnel. Here, e(t) = L d (t)-L rL r This represents the desired distance, which is usually set as a constant, i.e., the distance that the autonomous vehicle and the tracking personnel are expected to maintain.

[0043] PID control calculates the acceleration / deceleration signal acc(t) of the unmanned vehicle at the current time t to control its movement. Common PID parameters include proportional gain, integral gain, and derivative gain; however, in this embodiment 1, the PID parameter uses K′. P (t), K′ I (t), K′ D (t). Wherein, K′ P (t) is K P The corrected value obtained after processing by the DDPG network model; K′ I (t) is K I The corrected value obtained after processing by the DDPG network model; K′ D (t) is K D The corrected value obtained after processing by the DDPG network model. K P Preset value for proportional gain, K I Preset value for integral gain, K D This is the preset value for the differential gain.

[0044] Right now: In the formula, e(t-1) represents the error of the previous time t-1, and Δt represents the interval between the previous time t-1 and the current time t.

[0045] In other words, the vision-based human tracking method for unmanned vehicles disclosed in Embodiment 1 includes the following steps:

[0046] 1. Obtain the actual distance L between the unmanned vehicle and the target person at the current time t. d (t), the angle α(t) between the direction of the line connecting the rear axle of the unmanned vehicle to the target person at the current time t and the direction of the unmanned vehicle's movement;

[0047] For L d (t), its calculation formula is:

[0048]

[0049] Among them, X r (t), Y r (t) represents the x and y coordinates of the rear axle center of the unmanned vehicle at the current time t; X p (t), Y p (t) represent the x and y coordinates of the target person at the current time t. Generally, by installing an inertial navigation system on an unmanned vehicle, X can be directly obtained. r (t), Y r (t). X p (t), Yp (t) Obtained through a camera-based target tracking method: Specifically, using camera-based target tracking technology, video image sequences are analyzed, and the detected target personnel region is matched to locate X. p (t), Y p (t).

[0050] For α(t), the formula for calculation is:

[0051]

[0052] φ(t) represents the heading angle of the unmanned vehicle at the current time t, which can also be obtained directly from the inertial navigation system installed on the unmanned vehicle.

[0053] 2. Obtain the preset value K of the proportional gain. P Integral gain preset value K I Differential gain preset value K D ;

[0054] For K P K I K D The values ​​are generally determined based on experience or experiments.

[0055] 3. Construct the DDPG network model and analyze L d (t), K P K I K D Processing yields L d ′(t), K′ P (t), K′ I (t), K′ D (t); where L d ′(t) is L d The correction value of (t), K′ P (t) is K P Correction value, K′ I (t) is K I Correction value, K′ D (t) is K D The correction value.

[0056] The DDPG network model is based on continuous action reinforcement learning. The DDPG network model consists of a policy network and an evaluation network. The policy network includes an input layer, intermediate layer 1, intermediate layer 2, and an output layer. The input layer has 3 neurons; intermediate layer 1 and intermediate layer 2 each have 12 neurons with the ReLU activation function; and the output layer has 4 neurons with the tanh activation function. The evaluation network also includes an input layer, intermediate layer 1, intermediate layer 2, and an output layer. The input layer has 7 neurons; intermediate layer 1 and intermediate layer 2 each have 12 neurons with the ReLU activation function; and the output layer has 1 neuron.

[0057] In this context, the number of neurons in the input layer of the policy network corresponds to the number of state variables.

[0058] In this embodiment 1, the policy network unit inputs L d (t), which corresponds to the generation of 3 state variables L. d (t), e(t), and acc(t) are combined to form the state group s at the current time t. t =[L d (t), e(t), acc(t)].

[0059] Set the reward r at the current time t. t =-(e(t) 2 +acc(t) 2 );r t If r is negative, during the training process t The closer to 0, the better. In other words, term 'e' indicates that the actual distance between the autonomous vehicle and the person is as close as possible; term 'acc' indicates that the absolute value of the acceleration during the following process should be as small as possible.

[0060] Since the ReLU activation function is The tanh activation function is

[0061] Thus, the output range of the policy network is limited to [-1, 1], that is:

[0062]

[0063] In the formula, μ ddpg This represents the policy network, where subscripts 1, 2, 3, and 4 correspond to the four neurons in the output layer of the policy network. 's' represents the state, and 'θ' represents the output layer. μ Indicates the network parameters of the strategy.

[0064] Since the output range of the four neurons in the output layer of the policy network is [-1, 1], which is equivalent to L... d (t), K P (t), K I (t), K D (t) multiplied by a coefficient between [0,2], yields L. d ′(t), K P ′(t), K I ′(t), K D ′(t).

[0065] The number of input layer neurons in the evaluation network corresponds to the sum of the number of state variables and the number of action variables.

[0066] In this embodiment 1, the evaluation network input L d (t), e(t), acc(t), L d ′(t), K P ′(t), K I ′(t), K D ′(t), a total of 7; the evaluation network department guides the network department to update network parameters.

[0067] More specifically, the training method for the DDPG network model is as follows:

[0068] 1. Initialize the policy network μ ddpg (s|θ μ Initialize the policy network parameters θ μ Where s represents the state;

[0069] 2. Initialize the target network μ' ddpg (s|θ′ μ Initialize the target network parameters θ′ μ ;

[0070] 3. Initialize and evaluate the network Q ddpg (s, a|θ) Initialize and evaluate the network parameters θ; a represents the action;

[0071] 4. Initialize and evaluate the target network Q' ddpg (s, a|θ'), initialize and evaluate the target network parameters θ′;

[0072] 5. Initialize experience pool D;

[0073] 6. for episode = 1:M, where M represents the number of iteration rounds and episode represents a single iteration round;

[0074] 7. Initialize state s0;

[0075] 8. for t = 1:Tepi, where Tepi represents the total time of a single iteration round, and t represents the current time of a single iteration round;

[0076] 9. Based on the exploration noise, execute the action at the current time t. s t This represents the state group at time t. This represents the exploration noise at the current time t;

[0077] 10. After the agent performs an action, it reaches the next state group s. t+1 Obtain the corresponding reward r from the environment t ;r t This represents the reward at the current time t;

[0078] 11. Translate a single experience (s) t ,a t ,r t ,s t+1 Store it in experience pool D;

[0079] 12. Select m experiences (s) from the experience pool D. i ,a i ,r i ,s i+1 ), i∈[1,m]; where, s i Let a represent the i-th state variable in the experience pool. i r represents the i-th action quantity in the experience pool. i s represents the i-th reward in the experience pool. i+1 This represents the (i+1)th state variable in the experience pool.

[0080] 13. Experience reward based on selection r i and evaluation of target network computation This indicates that the policy network updates the target value, and γ represents the discount factor. This represents the estimate of the (i+1)th action in the policy network;

[0081] 14. Update and evaluate network parameters θ based on TD_error; TD_error represents the time difference error.

[0082] 15,

[0083] 16. Based on the evaluation network, calculate the deterministic policy gradient and update the policy network parameters θ. μ ;

[0084] 17, Represents θ μ Update gradient, J(θ) μ )express The loss value, Indicates a i Update gradients, Let μ(s) represent the estimate of the i-th action in the policy network. i |θ μ ) indicates that the policy network is based on s i The output value;

[0085] 18. Update the target network parameters θ′ of the policy. μ And evaluate the target network parameters θ′;

[0086] 19, θ′ μ =τθ′ μ +(1-τ)θ μ τ represents the maintenance coefficient;

[0087] 20, θ′=τθ′+(1-τ)θ;

[0088] 21, if done;

[0089] 22, break;

[0090] 23, end;

[0091] 24, end.

[0092] Combining the training methods described above, the DDPG network model undergoes M iterations (i.e., M training iterations), with each iteration adjusting L... d (t), K P K I K D After adjustment, we obtain L. d ′(t), K P ′(t), K I ′(t), K D The network parameters of the policy network are updated synchronously, and after M iterations of training, the network parameters of the policy network are fixed and no longer change.

[0093] In summary, the DDPG network model undergoes M iterations of training. After each iteration, the policy network outputs the corrected values ​​of the autonomous vehicle control parameters obtained in that iteration. After each iteration, the evaluation network updates the network parameters of the policy network. After M iterations, the parameters of the policy network are fixed.

[0094] It should be noted that, due to L d (t) is a variable, therefore the network department of the strategy with fixed network parameters is still dynamically adjusted. d ′(t), K P ′(t), K I ′(t), KD ′(t).

[0095] Based on the steps above, for driverless cars, we have:

[0096] The method of lateral control is as follows:

[0097] Calculate the front wheel angle of the unmanned vehicle at the current time t. And according to δ f (t) to make a turn; where L represents the wheelbase of the driverless car.

[0098] The method of longitudinal control is as follows:

[0099] Compare L d (t) and the preset expected distance L r The error at the current time t is obtained as e(t) = L. d (t)-L r ;

[0100] The unmanned vehicle's speed is controlled based on PID control according to e(t) to maintain a safe distance between the unmanned vehicle and the target personnel; the PID parameter uses K′ P (t), K′ I (t), K′ D (t).

[0101] Specifically, calculate the acceleration / deceleration signal acc(t) of the unmanned vehicle at the current time t;

[0102]

[0103] Where e(t-1) represents the error of the previous time t-1, and Δt represents the interval between the previous time t-1 and the current time t;

[0104] If e(t)≥0, and L d If acc(t) is not greater than the preset distance threshold, and acc(t) is positive, then the unmanned vehicle accelerates at the current time t according to the acceleration of acc(t).

[0105] If e(t) < 0, and L d If acc(t) is not less than the preset distance threshold of 2, and acc(t) is negative, then the unmanned vehicle decelerates according to the acceleration of acc(t) at the current time t.

[0106] If e(t)≥0 and e(t) is greater than the preset distance threshold, then the unmanned vehicle stops moving at the current time t.

[0107] If e(t) < 0, and L d If (t) is less than the preset distance threshold, then the unmanned vehicle stops moving at the current time t.

[0108] In this embodiment 1, the preset distance threshold is 2L. r That is, e(t) > 2L r This indicates that the distance between the autonomous vehicle and the target person is too far, which could easily cause obstruction and measurement errors, leading to failure of subsequent following. Therefore, in this case, the vehicle chooses to stop moving and cease following. The preset distance threshold is 1.5L, i.e., L d If (t) < 1.5L, it indicates that the front of the unmanned vehicle is too close to the target personnel, and it should stop moving to ensure safety.

[0109] Of course, the preset distance threshold one and preset distance threshold two can also be adjusted according to the actual situation.

[0110] Example 2

[0111] This embodiment 2 discloses an unmanned vehicle vision-based personnel following system, which uses the unmanned vehicle vision-based personnel following method of embodiment 1.

[0112] The vision-based human-following system for autonomous vehicles includes: a parameter acquisition module, a DDPG network module, a lateral control module, and a longitudinal control module.

[0113] The parameter acquisition module is used to obtain the actual distance measurement value L between the unmanned vehicle and the target person at the current time t. d (t), the angle α(t) between the line connecting the rear axle of the unmanned vehicle to the target person at the current time t and the forward direction of the unmanned vehicle; also used to obtain the preset value K of the proportional gain. P Integral gain preset value K I Differential gain preset value K D .

[0114] The DDPG network module is used to build DDPG network models and to handle L... d (t), K P K I K D Processing yields L d K′(t), K′X(t), K′ I (t), K′ D (t). Wherein, L d ′(t) is L d The correction value of (t), K′ P (t) is K P Correction value, K′ I (t) is K I Correction value, K′ D (t) is K D The correction value.

[0115] The lateral control module is used to calculate the front wheel steering angle of the unmanned vehicle at the current time t. And according to δ f (t) performs a turn. Where L represents the wheelbase of the autonomous vehicle.

[0116] The longitudinal control module is used to compare L d (t) and the preset expected distance L r The error at the current time t is obtained as e(t) = L. d (t)-L r The system controls the vehicle speed based on PID control according to e(t), ensuring a safe distance between the vehicle and the target personnel. The PID parameters are K′. P (t), K′ I (t), K′ D (t).

[0117] This embodiment 2 also discloses an unmanned vehicle that uses the vision-based human following method of the unmanned vehicle in embodiment 1.

[0118] Example 3

[0119] This embodiment 3 discloses a readable storage medium storing computer program instructions. When the computer program instructions are read and executed by a processor, the steps of the vision-based human following method for unmanned vehicles in embodiment 1 are performed.

[0120] When applying the method of Example 1, it can be applied in the form of software, such as a program designed to run independently on a computer-readable storage medium, which can be a USB flash drive or a USB security token, and designed to be a program that starts the entire method through an external trigger.

[0121] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0122] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A vision-based human tracking method for unmanned vehicles, used for lateral and longitudinal control of the unmanned vehicle, characterized in that, The vision-based human tracking method for autonomous vehicles includes the following steps: Get the current time t Actual distance measurement between the unmanned vehicle and the target person Current moment t The angle between the direction of the line connecting the rear axle of the autonomous vehicle to the target person and the direction of travel of the autonomous vehicle. ; Get the proportional gain preset value Integral gain preset value Differential gain preset value ; Construct a DDPG network model and perform... , , , Processed , , , ;in, for Correction value for Correction value for Correction value for Correction value; The method of lateral control is as follows: Calculate the current time t Front wheel angle of driverless car and in accordance with To make a turn; among them, L This indicates the wheelbase of the driverless car; The method of longitudinal control is as follows: Compare Distance from the preset target Get the current time t error ; Based on PID The speed of the autonomous vehicle is controlled to maintain a safe distance between the vehicle and the target personnel; the parameters of the PID controller are adopted. , , ; The calculation formula is: ； in, , Each represents the current time. t The horizontal and vertical coordinates of the rear axle center of the driverless car; , Each represents the current time. t The horizontal and vertical coordinates of the target personnel; The calculation formula is: ； in, Indicates the current time t The heading angle of the autonomous vehicle; The DDPG network model includes: a policy network unit and an evaluation network unit; the DDPG network model performs... M The training process involves multiple iterations; the policy network unit outputs corrected values ​​for the autonomous vehicle control parameters obtained in each iteration; the evaluation network unit updates the network parameters of the policy network unit after each iteration; and after... M After the next iteration of training, the network parameters of the policy network are fixed. The policy network unit includes an input layer, an intermediate layer one, an intermediate layer two, and an output layer; wherein, the input layer of the policy network unit has 3 neurons; the intermediate layer one of the policy network unit has 12 neurons, and the activation function of this layer is ReLU; the intermediate layer two of the policy network unit has 12 neurons, and the activation function of this layer is ReLU; the output layer of the policy network unit has 4 neurons, and the activation function of this layer is tanh. The evaluation network also includes an input layer, an intermediate layer one, an intermediate layer two, and an output layer; wherein, the input layer of the evaluation network has 7 neurons; the intermediate layer one of the policy network has 12 neurons, and the activation function of this layer is ReLU; the intermediate layer two of the policy network has 12 neurons, and the activation function of this layer is ReLU; the output layer of the policy network has 1 neuron.

2. The vision-based human tracking method for unmanned vehicles according to claim 1, characterized in that, , , Values ​​are determined based on experience or experimentation.

3. The vision-based human tracking method for unmanned vehicles according to claim 1, characterized in that, Based on PID Controlling the speed of autonomous vehicles includes: Calculate the current time t Autonomous vehicle acceleration and deceleration signals ; ； in, Indicates the previous moment t -1 error Indicates the previous moment t -1 and the current time t The interval; like ,and Not greater than a preset distance threshold of 1 If it is positive, then at the current moment t driverless cars according to Acceleration is achieved through acceleration; like ,and Not less than a preset distance threshold of two. If it is negative, then at the current moment t driverless cars according to The acceleration is used to decelerate; like ,and If the distance is greater than a preset distance threshold of 1, then at the current time... t The driverless car stopped moving; like ,and If the distance is less than the preset distance threshold of two, then at the current time... t The driverless car stopped moving.

4. The vision-based human tracking method for unmanned vehicles according to claim 3, characterized in that, The preset distance threshold is... The preset distance threshold is 2. .

5. A vision-based human-following system for unmanned vehicles, characterized in that, It uses the vision-based human following method for unmanned vehicles as described in any of claims 1-4; The driverless vehicle's vision-based human-following system includes: The parameter acquisition module is used to obtain the current time. t Actual distance measurement between the unmanned vehicle and the target person Current moment t The angle between the direction of the line connecting the rear axle of the autonomous vehicle to the target person and the direction of travel of the autonomous vehicle. It is also used to obtain the preset value of the proportional gain. Integral gain preset value Differential gain preset value ; The DDPG network module is used to build DDPG network models and... , , , Processed , , , ;in, for Correction value for Correction value for Correction value for Correction value; The horizontal control module is used to calculate the current time. t Front wheel angle of driverless car and in accordance with To make a turn; among them, L This indicates the wheelbase of the driverless car; as well as The longitudinal control module is used for comparison. Distance from the preset target Get the current time t error And based on PID The speed of the autonomous vehicle is controlled to maintain a safe distance between the vehicle and the target personnel; the parameters of the PID controller are adopted. , , .

6. An unmanned vehicle, characterized in that, It uses the vision-based human following method for unmanned vehicles as described in any of claims 1-4.

7. A readable storage medium, characterized in that, The readable storage medium stores computer program instructions, which are read and executed by a processor to perform the steps of the vision-based human following method for unmanned vehicles as described in any of claims 1-4.

Images