Autonomous target following control method fusing multiple judgment modes for robot

Abstract

The invention discloses a multi-judgment-mode-fused autonomous target following control method for a robot, and the method comprises the steps: enabling the robot to obtain the color and depth information of a surrounding environment through a carried depth camera, detecting whether a target object exists in a color image through a target detection algorithm, obtaining the pixel coordinates of the target object if the target object exists, and carrying out the recognition of the target object; meanwhile, distance information is acquired by using the depth image. And if the target is lost, predicting the position of the target by using Kalman filtering. And when the predicted time exceeds a threshold value, the UWB is used for repositioning the position and the distance of the target. And finally, calculating a linear velocity and an angular velocity of robot following according to the position information obtained in the above mode through a control strategy, and realizing autonomous following. According to the method, multiple judgment modes are fused, the accuracy of target recognition and judgment in the robot following process can be improved, and the method has the advantages of good following stability, low loss rate, low false alarm rate and the like and can be suitable for more diversified working scenes.

Description

Technical Field

[0001] The present invention relates to robot following control, and more specifically to an autonomous target following control method for a robot integrating multiple determination modes, and belongs to the technical field of robot control. Background Art

[0002] With the progress of the times, various cutting-edge high-tech technologies have undergone multidisciplinary cross-upgrades, and robots are a typical representative of them. Robots are known as "the pearl on the crown of manufacturing industry". Their research and development, manufacturing and application are important criteria for measuring a country's scientific and technological innovation and high-end manufacturing level. The robot's wheeled platform can perceive and interact with the surrounding environment through sensors and monitor its own operating status. It can follow the task target with an optimized following strategy to complete dynamic production tasks. The existing logistics centers and airports in cities need to transport goods. If manual handling is used, it is not only inefficient, but also has high operating costs and high labor intensity. The transport robot that can follow intelligently can solve these problems well. At present, the storage robots in many logistics centers can realize functions such as personnel following and obstacle avoidance. Airports use following robots to help passengers carry luggage, improve transportation efficiency and save transportation costs. Among them, wheeled platforms mainly include wheeled, crawler and legged platforms.

[0003] During the following driving process of a wheeled robot, the stability of the target being followed is easily disturbed by a variety of factors, such as changes in lighting, the sudden appearance of obstructions, the similarity of the shape and color of the target object, as well as sensor noise and data loss. These factors may lead to target recognition errors or losses, which in turn cause the deviation of the following path and the reduction of control accuracy. In addition, the dynamic changes in the environment and the increase in background complexity will also aggravate the perception error, affecting the robot's ability to continuously track the target and the response speed, thus causing instability in the following process.

[0004] The prior art discloses a following robot and a following control method and a following control system. The following control method includes: periodically obtaining the relative position between the following target and the following robot; judging whether the relative position exceeds the following range preset by the following robot, and if it does not exceed the following range, obtaining the output speed based on the relative position; judging whether there is an obstacle in the moving direction of the following robot, and if there is an obstacle, obtaining the obstacle avoidance linear speed, and using the obstacle avoidance linear speed as the driving speed of the following robot; and if there is no obstacle, using the output speed as the driving speed of the following robot.

[0005] Most existing technologies use a single determination strategy to identify targets. This method is susceptible to multiple interferences from environmental factors during dynamic following, such as illumination changes, background complexity, occlusion, and target morphology changes, resulting in insufficient robustness in target identification, which can lead to target loss and unstable following. In addition, a single determination method cannot effectively handle sensor data noise and uncertainty, further increasing the risk of misjudgment and failure of the following system. Summary of the invention

[0006] In view of the above-mentioned shortcomings of the following strategy in the prior art that adopts a single judgment method to identify the target, the purpose of the present invention is to provide an autonomous target following control method for a robot that integrates multiple judgment methods. This method integrates multiple judgment methods and can improve the accuracy of target recognition and judgment in the robot following process. It has the advantages of good following stability, low loss rate, low false alarm rate, etc. It can be applied to more diverse work scenarios, so that the robot can complete the following task more stably.

[0007] The technical solution of the present invention is achieved in this way:

[0008] An autonomous target following control method for a robot integrating multiple judgment modes is carried out in the following steps:

[0009] 1) The robot uses its own depth camera to obtain color images and depth images of the surrounding environment, and uses the target detection algorithm to detect whether there is a target object to be followed in the color image. If there is a target object, it proceeds to step 2); if there is no target object, it proceeds to step 3);

[0010] 2) Obtain target location information based on the image acquired by the depth camera and control the robot to follow the target;

[0011] 3) Use the Kalman filter algorithm to predict the target position, obtain the target position information through prediction and control the robot to follow the target;

[0012] While using the Kalman filter algorithm for prediction, determine whether the target object appears in the color image of the depth camera; if the target object appears, return to step 2), otherwise repeat step 3);

[0013] If the duration of continuous prediction using the Kalman filter algorithm exceeds the set value, proceed to step 4);

[0014] 4) Use UWB to relocate the target position, obtain the target position information based on the relocation, and control the robot to follow the target;

[0015] While using UWB to relocate the target position, determine whether the target object appears in the color image of the depth camera; if the target object appears, return to step 2), otherwise repeat step 4).

[0016] Furthermore, the target position information includes pixel coordinates and distance information; the linear velocity of the robot following is determined based on the distance information, and the angular velocity of the robot following is determined based on the pixel coordinates, and the robot is controlled to move according to the determined linear velocity and angular velocity, thereby achieving autonomous following of the target.

[0017] Furthermore, in step 2), when obtaining the target position information based on the image acquired by the depth camera, the pixel coordinates are determined by the color image of the depth camera, and the distance information is determined by the depth image of the depth camera.

[0018] The target detection algorithm is YOLOV5s.

[0019] Furthermore, MobileNetV3 is used to replace CSPDarknet in YOLOv5s as the backbone structure of the target detection network to form an MG-YOLOv5s lightweight neural network model, which can detect whether there is a following target in the color image and calculate the target position information.

[0020] Furthermore, in step 3), the target position at the previous moment is used to predict the target position at the current moment; if there is a target object in the color image of the depth camera at the previous moment, the target position collected by the depth camera at the previous moment is used as the target position at the previous moment, otherwise the target position at the previous moment is predicted by the Kalman filter algorithm.

[0021] Furthermore, in step 4), the method of relocating the target position using UWB to obtain the target position information is:

[0022] Two UWB devices are installed at the same height on the robot as base stations, and the lateral distance between the two base stations is greater than 0.5m. A UWB device is installed on the target object to be followed as a positioning bracelet; the positioning bracelet can transmit signals to achieve communication with the base station; the base station determines the distance from the positioning bracelet based on the received reflected signal, and the distance between the two base stations is predetermined, so the line between the two base stations and the positioning bracelet forms a triangle with determined three-side lengths, and the vertical distance and angle between the positioning bracelet and the line connecting the two base stations can be solved; the position of the positioning bracelet is the target position, thereby obtaining the target position information.

[0023] Furthermore, when controlling the robot to follow the target, the linear velocity and angular velocity are published through the ROS topic, and finally the VCU receives the command. The VCU uses the kinematic model to solve the linear velocity and angular velocity into the rotation speed of the robot's four wheel hub motors, and sends it to the motor controller, which controls the motor to rotate at the calculated speed.

[0024] Furthermore, the angular velocity ω and linear velocity speed of the robot are determined as follows:

[0025]

[0026]

[0027] Among them, ω max is the maximum angular velocity, w is a constant parameter, Δy is the offset of the target center; it is determined based on the difference between the pixel coordinates before and after; x is the distance between the robot and the target.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention uses a variety of methods such as depth camera, Kalman filter algorithm, UWB, etc. to identify and determine the follow-up target. By fusing information from multiple sensors or algorithms, the accuracy of target identification and determination can be improved. It has the advantages of good following stability, low loss rate, low false alarm rate, etc., and can be applied to more diverse work scenarios. In addition, the use cost of the above three methods is low and the benefits generated are high. The autonomous target following control strategy of the present invention has better robustness and real-time performance by fusing multiple determination methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is the MG-YOLOv5 model structure diagram.

[0031] Figure 2 This is a diagram showing the training results of MG-YOLOv5 on a self-made dataset.

[0032] Figure 3 This is the MG-YOLOv5 target detection flowchart.

[0033] Figure 4 This is a schematic diagram of UWB positioning.

[0034] Figure 5 This is a flow chart of the autonomous target following control strategy integrating multiple judgment methods of the present invention. DETAILED DESCRIPTION

[0035] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] The present invention is a method for autonomous target following control of a robot integrating multiple determination modes, and the steps are as follows:

[0037] 1) The robot uses its own depth camera to obtain color images and depth images of the surrounding environment, and uses the target detection algorithm to detect whether there is a target object to be followed in the color image. If there is a target object, it proceeds to step 2); if there is no target object, it proceeds to step 3);

[0038] 2) Obtain target location information based on the image acquired by the depth camera and control the robot to follow the target;

[0039] 3) Use the Kalman filter algorithm to predict the target position, obtain the target position information through prediction and control the robot to follow the target;

[0040] While using the Kalman filter algorithm for prediction, determine whether the target object appears in the color image of the depth camera; if the target object appears, return to step 2), otherwise repeat step 3);

[0041] If the duration of continuous prediction using the Kalman filter algorithm exceeds the set value, the prediction is considered to have failed and the process goes to step 4);

[0042] 4) Use UWB to relocate the target position, obtain the target position information based on the relocation, and control the robot to follow the target;

[0043] While using UWB to relocate the target position, determine whether the target object appears in the color image of the depth camera; if the target object appears, return to step 2), otherwise repeat step 4).

[0044] The target position information includes pixel coordinates and distance information; the linear velocity of the robot following is determined based on the distance information, and the angular velocity of the robot following is determined based on the pixel coordinates, and the robot is controlled to move according to the determined linear velocity and angular velocity, so as to achieve autonomous following of the target.

[0045] In order to facilitate robot following, a target frame is pre-set on the target object, and the robot follows the target object by following the target frame. At the same time, whether the target frame is detected is used to determine whether the target is lost. If there is no information about the target frame in the image obtained by the robot's depth camera, the target is considered lost, otherwise the target is not lost.

[0046] The present invention uses a variety of methods such as a depth camera, a Kalman filter algorithm, and UWB to identify and determine the follow-up target. By fusing information from a variety of sensors or algorithms, the accuracy of target identification and determination can be improved. It has the advantages of good follow-up stability, low loss rate, and low false alarm rate, and can be applied to more diverse work scenarios. The present invention preferentially uses images obtained by a depth camera to achieve target tracking and positioning. If the depth camera does not obtain target information, the target information is obtained by prediction; if the target is not found after the prediction time exceeds the threshold, it is considered that the prediction has failed, and finally the target position is relocated by UWB. Since the ranging range of UWB is far, but the signal of UWB diverges outdoors, its ranging accuracy is poor. If UWB is used for control at a closer distance, its distance information will always fluctuate greatly, and the control is not accurate. The distance information measured by the depth camera within 5 meters is more accurate and has less fluctuations, because the distance followed by the robot is usually controlled at about 5 meters, and in most cases within 5 meters. Therefore, the present invention mainly and primarily uses the information of the depth camera for control, and the other two positioning methods are supplementary, so that the control will be more accurate. The Kalman filter algorithm makes predictions based on the information measured by the depth camera. Therefore, the Kalman filter algorithm can only be used for prediction after the depth camera has information. If the target can be found under the prediction of the Kalman filter algorithm, the information measured by the depth camera is still used for control after it is retrieved. The UWB has a large ranging range. When a person walks very far, the marker appears very small, and the target detection algorithm may miss detections, misdetections, or fail to detect targets that are too small. Therefore, the use of UWB positioning in the end just gives full play to the advantages of UWB. Therefore, the present invention locates the target according to the three priorities of depth camera, Kalman filter algorithm, and UWB positioning, which can maximize the advantages of each, achieve reliable tracking of the target, and be able to retrieve the target as soon as possible after it is lost.

[0047] The present invention determines whether there is a target object to be followed by the color image of the depth camera. If there is a target object, the pixel coordinates are further determined by the color image, and the distance information is determined by the depth image. When determining specifically, the target detection algorithm is used to detect whether there is a target object to be followed in the color image and calculate the target position information. To this end, the present invention proposes a lightweight MG-YOLOv5s target detection model, which is as follows.

[0048] The basis of the MG-YOLOv5s target detection model is YOLOv5s. The YOLOv5s model is mainly divided into three parts, namely the backbone network, the neck network and the decoupling head. The backbone network of the YOLOv5s model is mainly composed of the Focus and CSP structures. The Focus structure is a convolutional layer used for feature extraction in neural networks. It compresses and combines the information in the input feature map to extract higher-level feature representations. The Focus structure divides the input feature map into four sub-maps and splices the sub-maps through channels to obtain a smaller feature map. The CSP structure divides the input feature map into two parts: one part is processed by a small convolutional neural network, and the other part directly enters the next layer for processing. Finally, the two parts of the feature map are spliced ​​together as the input of the next layer. The CSP structure significantly reduces the number of parameters and calculations of the neural network, and improves the reasoning and training efficiency of the model. The neck network of YOLOv5s is mainly composed of the SPP (Spatial Pyramid Pooling) and PAN (Path Aggregation Network) structures. The SPP structure is a feature pyramid pooling structure that uses pooling layers of different sizes to pool feature maps and fuse information of different scales. The PAN structure is mainly composed of a feature pyramid module and a feature fusion module. Through multi-level feature fusion, the model's detection ability for objects of different scales is improved.

[0049] In order to further reduce the amount of parameter calculation of the convolutional neural network model, enable the model to be deployed on the edge data processing platform, reduce the computational complexity and improve the efficiency of feature extraction, the present invention uses MobileNetV3 to replace the CSPDarknet in YOLOv5s as the backbone structure of the target detection network, forming the MG-YOLOv5s lightweight neural network model. This network has the characteristics of high detection accuracy, light weight, fast detection speed and low latency. The introduction of DWConv and C3Ghost modules in the Neck layer makes the model more lightweight, effectively reducing the number of parameters and calculation of the model without losing detection accuracy, making it suitable for mobile devices or resource-constrained environments while maintaining good performance. Finally, an improved MG-YOLOv5 network architecture is formed, such as Figure 1 As shown in the figure, the original YOLOv5s model has 7225885 parameters and 16.6GFLOPs of computation, while the improved MG-YOLOv5s model has 2050246 parameters and 3.6GFLOPs of computation.

[0050] More than 10,000 images of self-made markers were collected, covering indoor and outdoor, different scenes, angles and distances, etc., to form a self-made data set for training. The trend of each indicator during the training process is as follows Figure 2As shown. In the training and validation datasets, positioning, confidence, and classification losses are key indicators that reflect the difference between the model's predicted value and the true value. Since only a single category is involved, the classification loss can be ignored. As the number of training times increases, these losses gradually decrease and stabilize, indicating that the model performance has improved and is more stable. In addition, accuracy, recall, and two mean average precision (mAP) indicators are used to evaluate the performance of the model in classification problems. They measure the detection accuracy of the model, and the higher the value, the better the model performance. Because the training images are relatively simple and the homemade dataset is sufficient, the mAP50 accuracy can be as high as 99.5% after the evaluation indicator values ​​are stabilized.

[0051] The output end mainly outputs the prediction box information, which includes: confidence, category probability, and bounding box position. The object detection process based on the depth camera is as follows: Figure 3 shown.

[0052] When the target object is obscured or lost, in order to continuously obtain the target's position and distance information, the present invention adopts a Kalman filter algorithm to predict the position and distance of the lost target, that is, the target position at the previous moment is used to predict the target position at the current moment; if the target object exists in the color image of the depth camera at the previous moment, the target position collected by the depth camera at the previous moment is used as the target position at the previous moment, otherwise the target position at the previous moment is predicted by the Kalman filter algorithm.

[0053] The update frequency of the following robot command is high, and the robot motion between two frames is approximately considered to be a linear change, so the discrete time model of the robot's position in the image plane and the distance between the robot and the target is as follows:

[0054]

[0055] In the formula, x and y are the coordinates of the center point of the target frame, w and h are the width and height of the target frame, and l is the distance between the wheeled differential robot and the target. The robot's state quantity S i =[x i ,y i ,w i ,h i ,l i ,Δx i ,Δy i ,Δl i ], i = 0, 1, 2, where S0 is the coordinate of the first frame where the target is detected, P0 = (x0, y0), the detection box size D0 = (w0, h0), and the state estimation of the robot at the i-th moment is as follows:

[0056]

[0057] In the formula is the state estimate of the i-th frame, is the optimal state estimate of the i-1th frame, and F is the state transfer matrix, which is expressed as follows:

[0058]

[0059] The state estimation covariance matrix of the i-th frame is as follows:

[0060] P(i|i-1)=FP(i-1|i-1)F T +Q Formula 4

[0061] Where P(i-1|i-1) is the covariance matrix of the optimal estimate of the i-1th frame. In the target tracking process, the system process noise mainly comes from the uncertainty of the movement during the target movement. However, in the present invention, because the control instruction update frequency is high, it is assumed that the change between two frames during the movement of the target is approximately linear. The system process noise covariance matrix Q is as follows:

[0062]

[0063] For the Kalman gain K(i|i) of the i-th frame, the optimal estimate is The covariance matrix P(i|i) is updated:

[0064]

[0065] I is the identity matrix, Z(i|i) refers to the observation value in the Kalman filter;

[0066] Where H is the state transfer matrix, expressed as follows:

[0067]

[0068] Where R is the observation noise covariance matrix. In the target following process, H comes from target loss, overlap, and environmental changes, that is, only part of the target can be detected or no target can be detected. In a complex environment, the error of the system mainly comes from the observation noise. The noise in the actual environment often does not satisfy the Gaussian distribution. The new information adaptive Kalman filter is used to adaptively adjust the filtering process according to the actual observation noise.

[0069] The new information obtained by the visual sensor in the i-th frame is defined as the actual value Y(i|i) and the estimated value The difference between them is as follows:

[0070]

[0071]

[0072] In formula 9, vi For the actual value observation noise, substitute Equation 9 into Equation 8 and take the variance of both sides to get Equation 10:

[0073] cov[φ(i|i)φ T (i|i)]=HP(i|i-1)H T +R Formula 10

[0074] Combined with the sliding window method, the covariance of the new information estimate is as follows:

[0075]

[0076] Where m is the sliding window length. Combining equations 10 and 11, we get the observation noise covariance matrix of the i-th frame:

[0077]

[0078] When the target is detected, the IOU value between the detection frame of the i-1th frame and the i-th frame is calculated:

[0079]

[0080] S i is the area of ​​the detection box in the i-th frame. If the IOU value is greater than or equal to the threshold σ, the detection result of the i-th frame is considered to be the observation value of the system, and the state is updated The coordinate change of the center point of the parameter is (Δx(i|i-1), Δy(i|i-1)), and the distance change is Δl(i|i-1):

[0081]

[0082] In formula (14), Z(x(i|i)), Z(y(i|i)), and Z(l(i|i)) are the coordinates and distance of the center point of the detection box in the system observation parameters in the i-th frame, respectively.

[0083] When the target is lost or the IOU value is less than the threshold σ, because the system lacks observations, the optimal estimate of the previous frame is directly used to update the optimal estimate of the current frame, so AKF updates the system state The status at this time is:

[0084]

[0085] After the target is lost, if AKF fails to successfully predict the target position within a certain time range, it is considered that the AKF prediction function has failed, and UWB is used for repositioning. In order to realize the following function of the following robot, it is necessary to obtain the distance and lateral coordinates between the target and the robot. The TOA and TDOA positioning methods require at least three base stations and require the base stations to be fixed, which limits the use scenarios and is not suitable for the following requirements of the present invention. The UWB used by this robot cannot directly measure the angle between the tag and the base station, so the DOA positioning method cannot be used directly. Based on this, the present invention proposes a new UWB positioning method.

[0086] The UWB positioning of the present invention requires three UWB devices, two of which are placed in parallel at both ends of the robot as base stations, and the lateral distance between the two is guaranteed to be greater than 0.5m. The third UWB device is placed on the follow-up target, such as in the form of a positioning bracelet or tag. The positioning bracelet can transmit signals to achieve communication with the base station; the base station determines the distance from the positioning bracelet based on the received signal. Since the distance between the two base stations is fixed and determined in advance, the line between the two base stations and the positioning bracelet forms a triangle with a certain length of three sides. The vertical distance and angle between the positioning bracelet and the line connecting the two base stations can be solved through geometric relationships; the position of the positioning bracelet is the target position, and the target position information is obtained. According to the position of the target object, there are Figure 4 Two situations are shown.

[0087] exist Figure 4 In the figure, L1 and L2 are the straight-line distances of the target object from each base station, which are directly obtained by UWB, L is the distance between the two base stations, d is the vertical distance between the positioning bracelet and the line connecting the two base stations, and F is the offset of the target object from the center point of the two base stations. The figure is divided into two cases, a and b. The positioning of these two cases can be calculated by equations 16 and 17 respectively.

[0088]

[0089]

[0090] The robot uses the onboard depth camera to obtain the color information and depth information of the surrounding environment, and uses the target detection algorithm MG-YOLOV5s to detect whether there is a target object in the color image. If there is a target object, its pixel coordinates are obtained, and the corresponding distance information is read using the depth image. If the target is blocked or lost, the new information adaptive Kalman filter is used to predict the position and distance of the target. Because it is difficult to apply to long-term series prediction, when the prediction time exceeds a certain threshold, the prediction is judged to have failed. At this time, UWB is used to relocate the position and distance of the target. Finally, the position information-pixel coordinates and distance information obtained in the above manner are calculated through the control strategy to calculate the linear velocity and angular velocity of the platform follow, and the linear velocity and angular velocity are published through the ROS topic. Finally, the VCU receives the instruction, and the VCU uses the kinematic model to solve the linear velocity and angular velocity into the rotation speed of the four hub motors, and sends it to the motor controller, which controls the motor rotation. The overall process of the autonomous target following control strategy integrating multiple judgment methods of the present invention is as follows Figure 5 shown.

[0091] After obtaining the position information of the target object to be followed according to various judgment methods, if it is within the safe distance between people and vehicles, the robot speed is zero; when the distance is less than the ideal distance, the robot follows by gradually increasing the speed; when the distance is within the ideal distance range, the robot follows at the ideal speed; when the distance exceeds the ideal distance range, the robot follows by rapidly increasing the speed; finally, when the distance between the target and the robot exceeds the threshold, because the distance is large, there may be obstacles between the two. If the robot continues to follow, there will be safety hazards. Therefore, in this case, the robot speed is zero and stops to wait for the target to be re-identified. The angular velocity control function is shown in Equation 18, and the linear velocity control function is shown in Equation 19.

[0092]

[0093]

[0094] In equation 18, ω is the desired angular velocity, ω max is the maximum angular velocity, w is a constant parameter, Δy is the offset of the target center; if it is UWB positioning, it is F in equations 16 and 17; in equation 19, x is the distance between the robot and the target (in meters), if it is UWB positioning, it is d in equations 16 and 17. speed is the output linear velocity (meters per second).

[0095] The linear velocity calculation formula proposed in this application has a more radical output, that is, the slope of the curve is larger, but the following process is still very safe and no safety accidents will occur due to excessive speed. In this way, the target can be followed more closely and is not easy to lose.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the applicant has described the present invention in detail with reference to the preferred embodiments, ordinary technicians in the field should understand that modifications or equivalent substitutions of the technical solution of the present invention without departing from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. An autonomous target following control method for a robot integrating multiple determination modes, characterized in that: Follow the steps below, 1) The robot uses its own depth camera to obtain color images and depth images of the surrounding environment, and uses the target detection algorithm to detect whether there is a target object to be followed in the color image. If there is a target object, it proceeds to step 2); if there is no target object, it proceeds to step 3); 2) Obtain target location information based on the image acquired by the depth camera and control the robot to follow the target; 3) Use the Kalman filter algorithm to predict the target position, obtain the target position information through prediction and control the robot to follow the target; While using the Kalman filter algorithm for prediction, determine whether the target object appears in the color image of the depth camera; if the target object appears, return to step 2), otherwise repeat step 3); If the duration of continuous prediction using the Kalman filter algorithm exceeds the set value, proceed to step 4); 4) Use UWB to relocate the target position, obtain the target position information based on the relocation, and control the robot to follow the target; While using UWB to relocate the target position, determine whether the target object appears in the color image of the depth camera; if the target object appears, return to step 2), otherwise repeat step 4).

2. The autonomous target following control method for a robot integrating multiple determination modes according to claim 1, characterized in that: The target position information includes pixel coordinates and distance information; the linear velocity of the robot following is determined based on the distance information, and the angular velocity of the robot following is determined based on the pixel coordinates, and the robot is controlled to move according to the determined linear velocity and angular velocity, so as to achieve autonomous following of the target.

3. The autonomous target following control method for a robot integrating multiple determination modes according to claim 1, characterized in that: In step 2), when obtaining the target position information based on the image acquired by the depth camera, the pixel coordinates are determined by the color image of the depth camera, and the distance information is determined by the depth image of the depth camera.

4. The autonomous target following control method for a robot integrating multiple determination modes according to claim 1, characterized in that: The target detection algorithm is YOLOV5s.

5. The autonomous target following control method for a robot integrating multiple determination modes according to claim 4, characterized in that: MobileNetV3 is used to replace CSPDarknet in YOLOv5s as the backbone structure of the target detection network to form the MG-YOLOv5s lightweight neural network model, which can detect whether there is a following target in the color image and calculate the target position information.

6. The autonomous target following control method for a robot integrating multiple determination modes according to claim 1, characterized in that: Step 3) When predicting, the target position at the previous moment is used to predict the target position at the current moment; if there is a target object in the color image of the depth camera at the previous moment, the target position collected by the depth camera at the previous moment is used as the target position at the previous moment, otherwise the target position at the previous moment is predicted by the Kalman filter algorithm.

7. The autonomous target following control method for a robot integrating multiple determination modes according to claim 2, characterized in that: Step 4) The method of using UWB to relocate the target position and obtain the target position information is as follows: Two UWB devices are installed at the same height on the robot as base stations, and the lateral distance between the two base stations is greater than 0.5m. A UWB device is installed on the target object to be followed as a positioning bracelet; the positioning bracelet can transmit signals to achieve communication with the base station; the base station determines the distance from the positioning bracelet based on the received reflected signal, and the distance between the two base stations is predetermined, so the line between the two base stations and the positioning bracelet forms a triangle with determined three-side lengths, and the vertical distance and angle between the positioning bracelet and the line connecting the two base stations can be solved; the position of the positioning bracelet is the target position, thereby obtaining the target position information.

8. The autonomous target following control method for a robot integrating multiple determination modes according to claim 2, characterized in that: When controlling the robot to follow a target, the linear velocity and angular velocity are published through the ROS topic, and finally the VCU receives the command. The VCU uses the kinematic model to solve the linear velocity and angular velocity into the rotation speed of the robot's four wheel hub motors, and sends it to the motor controller, which controls the motor to rotate at the calculated speed.

9. The autonomous target following control method for a robot integrating multiple determination modes according to claim 2, characterized in that: Determine the robot's angular velocity ω and linear velocity speed by the following formula: Among them, ω max is the maximum angular velocity, w is a constant parameter, Δy is the offset of the target center; it is determined based on the difference between the pixel coordinates before and after; x is the distance between the robot and the target.

Images