A multi-robot cooperative target recognition and tracking method in dynamic environment

Abstract

The application discloses a kind of dynamic environment under multi-robot cooperative target identification and tracking method, comprising: real-time identification target personnel;If the confidence of target personnel is lower than threshold and there is effective detection frame, then start re-identification module, identity confirmation is carried out using TCN network in combination with timing feature and motion pattern;If target is lost, start multi-view cooperative mechanism, by multi-view image splicing and distributed target fusion technology, construct global panoramic perception view, simultaneously, using the complementary characteristics of multi-robot view, solve the problem of tracking interruption caused by target occlusion;If the confidence of target personnel is higher than threshold, then real-time target position is acquired, and the closed-loop system from target identification to motion control is established using optimized APF and PID control strategy, and the tracking strategy is maintained.The application effectively overcomes the problem of visual obstruction, target loss and dynamic interference in complex environment, and realizes high-precision and high-stability target tracking.

Description

Technical Field

[0001] This invention belongs to the field of multi-robot collaborative perception and intelligent decision-making, specifically relating to a method for multi-robot collaborative target recognition and tracking in dynamic environments. Background Technology

[0002] Currently, target tracking technology based on multi-sensor fusion has made significant progress. Researchers mainly use technologies such as LiDAR, visual cameras, and ultra-wideband (UWB) to achieve target detection and localization, and have developed corresponding robot navigation and control algorithms based on this. However, LiDAR generates sparse point cloud data, making it difficult to capture detailed target information such as appearance and texture. Furthermore, processing the point cloud data captured by LiDAR requires significant computational resources in target tracking tasks. UWB typically requires the target to carry a tag (such as a UWB beacon), which is limited by equipment. Additionally, the localization accuracy of UWB is affected by environmental multipath effects (signal reflection), and its accuracy may decrease in complex environments. However, cameras have advantages such as small size, low cost, the ability to capture high-resolution images, and the provision of rich texture and color information, making them suitable for target recognition and tracking tasks. Therefore, research on autonomous tracking of designated personnel based on visual sensors is on the rise.

[0003] For multi-robot collaborative target recognition and tracking control tasks, researchers have adopted various methods. For example, Koide et al. developed the hdl_graph_slam system for hospital navigation, but this is limited by the requirement for pre-built mapping. Tasaki et al. fused LiDAR and IMU perception and verified the flexibility of this approach. Pang et al. further integrated visual and LiDAR data, using a combination of A* and TEB algorithms to complete target following; however, this method suffers from insufficient real-time performance due to the need for complete path planning. Sun and Jun et al. employed a dual-modal control strategy, dividing the following process into two independent modules: target tracking and obstacle avoidance, ensuring safe following through an intelligent switching mechanism. However, the motion discontinuity caused by speed adjustment during this mode switching process still needs further optimization.

[0004] This invention addresses key technical challenges in multi-robot cooperative target recognition and tracking control under dynamic environments, and constructs a complete solution integrating lightweight detection, robust tracking, and intelligent cooperative control. It demonstrates significant advantages in detection accuracy, tracking robustness, and cooperative control efficiency. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a multi-robot collaborative target recognition and tracking method in dynamic environments. Through multi-robot collaborative perception and intelligent decision-making, it effectively overcomes problems such as field of view obstruction, target loss, and dynamic interference in complex environments, achieving high-precision and high-stability target tracking. It can provide decision-making support for multi-robot target recognition and tracking in dynamic environments.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for multi-robot cooperative target recognition and tracking in a dynamic environment, the method comprising:

[0008] Real-time identification of target personnel;

[0009] If the confidence level of the target person is lower than the threshold and there is a valid detection box, the re-identification module is activated, and the TCN network is used to combine temporal features and motion patterns to confirm the identity.

[0010] If the target is lost, a multi-view collaboration mechanism is activated. Through multi-view image stitching and distributed target fusion technology, a global panoramic perception view is constructed. At the same time, the complementary characteristics of multiple robot perspectives are utilized to solve the tracking interruption problem caused by target occlusion.

[0011] If the confidence level of the target person is higher than the threshold, the target position is acquired in real time, and an optimized APF and PID control strategy is used to establish a closed-loop system from target recognition to motion control to maintain the tracking strategy.

[0012] Preferred methods for real-time identification of target personnel include:

[0013] Based on the YOLOv5s model, the task of real-time identification of target personnel is achieved;

[0014] The YOLOv5s model uses MobileNetV3 to reconstruct the backbone network, integrates the CA attention mechanism and BiFPN structure in the neck network, and optimizes the detection head performance by introducing the EIOU loss function.

[0015] Preferably, if the confidence level of the target person is below a threshold and a valid detection box exists, the re-identification module is activated. The method of identity verification using a TCN network combined with temporal features and motion patterns includes:

[0016] Predicted output: Network input It consists of the positional change of the target bounding box between adjacent time steps, specifically expressed as the geometric difference between the detection box at the current time step and the previous time step:

[0017] ;

[0018] The network input includes the target's horizontal and vertical positional offsets in the image coordinate system, as well as the changes in the length and width of the detection box. Through layer-by-layer feature extraction via temporal convolutional layers, the network ultimately outputs the predicted position and scale changes of the bounding box for future multi-step distances, forming a complete motion trajectory prediction result.

[0019] in, Let be the Y-axis coordinate at time t. Let be the X-axis coordinate at time t. Let t be the measurement frame width. Let t be the height of the detection frame. and They are respectively The changes of the target along the X and Y axes in the image coordinate system within a given time period. and They are respectively The changes in the width and height of the bounding box within a given time step are analyzed through layer-by-layer feature extraction by the TCN. The final output is the predicted position of the bounding box at the preset step size, along with the scale change, forming a complete motion trajectory prediction result, denoted as... :

[0020] ;

[0021] in, The step size for network prediction. for The predicted trajectory at any given moment;

[0022] Event triggering mechanism: Under normal tracking conditions, the system ignores the prediction output of the temporal network, maintains the basic classification process, and keeps a fixed length. Timing sliding window When a target loss event is detected, a re-identification verification mechanism is activated, cross-validating the motion trajectory prediction result with the appearance classification output; among which,

[0023] ;

[0024] in, for The predicted trajectory at any given moment;

[0025] Overlap rate calculation: A credibility score index is introduced into the re-identification verification process. and The calculation formula is as follows:

[0026] ;

[0027] ;

[0028] By comparing the bounding boxes output by the detection network With re-identification prediction boxes The degree of overlap is considered in conjunction with a preset threshold for decision-making. When the overall credibility assessment result exceeds the set threshold, the target re-identification is deemed successful, and the predicted bounding box compensates for the detected missing region; otherwise, it is considered a matching failure.

[0029] .

[0030] Preferably, if the target is lost, a multi-view collaborative mechanism is activated. This mechanism uses multi-view image stitching and distributed target fusion technology to construct a global panoramic perception view. Simultaneously, leveraging the complementary characteristics of multiple robot perspectives, methods to address tracking interruptions caused by target occlusion include:

[0031] A color transfer algorithm is introduced in the image preprocessing stage to establish a color mapping relationship between the source image and the target image;

[0032] The original image region S is recursively divided into four equal parts, and the center point coordinates of each sub-block are located in each division process;

[0033] Under the condition of a fixed number of neighborhood points N, the maximum threshold that allows a pixel point P to still be identified as a corner point is defined as the corner response value of that point.

[0034] The Harris response value is introduced as an auxiliary evaluation index. The FAST algorithm is used to characterize the corner intensity through the offset vector. The ORB algorithm is used to determine the corner direction through the intensity centroid. The concept of scale pyramid is introduced and implemented by generating images at multiple scale levels.

[0035] After obtaining the improved ORB feature points, FLANN is used for nearest neighbor matching, and the PROSAC algorithm is used to remove mismatched points.

[0036] Preferably, after obtaining the improved ORB feature points, the method of using FLANN for nearest neighbor matching and combining it with the PROSAC algorithm to remove mismatched points includes:

[0037] The set of feature point pairs obtained from FLANN coarse matching According to the matching quality function Sort in descending order to generate an ordered set of feature point pairs. ;

[0038] from The first n high-quality matching points are selected sequentially to form a set N, and the quality standard of set N is defined.

[0039] The feature point pairs are divided into groups of 4 points in descending order of quality. The sum of the quality of each group is calculated and sorted. The 4 groups of matching points with the highest quality are selected to calculate the homography matrix H.

[0040] The projection error of the remaining points is calculated using the homography matrix H;

[0041] The matching quality is evaluated by calculating the geometric error between the feature point and the corresponding projection point. If the error is less than the threshold, the feature point is regarded as an interior point; otherwise, it is regarded as an exterior point.

[0042] Compare the number of inliers with a preset threshold. If the condition is met, return the optimal model parameters; otherwise, continue iterating until the maximum number of iterations is reached.

[0043] Preferably, if the confidence level of the target person is higher than a threshold, the target position is acquired in real time, and an optimized APF and PID control strategy is used to establish a closed-loop system from target recognition to motion control. The methods for maintaining the tracking strategy include:

[0044] After confirming the target location, the robot calculates its trajectory based on the distributed APF. The gravitational component guides the robot to converge toward the target, while the repulsive component avoids obstacles.

[0045] A hierarchical PID architecture is adopted. The outer loop posture controller converts the desired coordinates into speed commands, and the inner loop speed controller tracks the target through encoder feedback and generates differential drive signals.

[0046] Preferably, after confirming the target location, the robot calculates its trajectory based on a distributed APF (Automatic Perception Function), with the gravitational component guiding the robot to converge toward the target and the repulsive component avoiding obstacles. The methods include:

[0047] Introducing random forces This provides additional power when the robot is stuck; the expression for this stochastic force is:

[0048] ;

[0049] in: It is the intensity coefficient of random force, used to control the magnitude of random disturbances; It is a random vector, with the direction in Uniformly distributed within a range, with sizes ranging from Randomly generated within the range;

[0050] After introducing random forces, the total resultant force acting on the robot can be expressed as the vector sum of attractive force, repulsive force, and random force:

[0051] ;

[0052] in, Represents gravity. Indicates repulsive force;

[0053] When the robot, the target point, and the obstacle are collinear, and the target point is centered, the robot will stop at a preset equilibrium point, where the net force is 0. The repulsive potential field function is then improved to make the repulsive potential energy at the target point zero. The repulsive potential field function is:

[0054] ;

[0055] in, Indicates the order, ; This indicates that the target person is located in a repulsive potential field; It is the repulsive force gain coefficient; It's the robot's location; It indicates the location of the obstacle; It is the distance between the robot and the obstacle; It is the furthest distance affected by the repulsive potential field of the obstacle; Indicates the target location; Indicates the distance between the robot and the target;

[0056] The repulsive force acting on the robot after the improved repulsive potential field function is:

[0057] ;

[0058] in, This represents the total repulsive force vector acting on the robot. The gradient represents the repulsive potential field. Represents the direct repulsive force of the obstacle. This represents the reverse repulsive force associated with the corresponding target point.

[0059] Preferably, a hierarchical PID architecture is adopted, in which the outer-loop pose controller converts the desired coordinates into velocity commands, and the inner-loop speed controller tracks the target through encoder feedback to generate differential drive signals. The method includes:

[0060] In the position control channel, the outer loop PID controller follows the desired distance. Distance from actual distance deviation Generate target linear velocity The inner-loop PID controller is responsible for... Track;

[0061] In the directional control channel, the outer loop PID controller operates based on the desired azimuth angle. Compared with actual azimuth deviation Calculate the target angular velocity And implemented by inner-loop PID. Rapid response;

[0062] The outputs of the two control channels are synthesized using a differential kinematics model: , ,in, The linear velocity of the robot is obtained from the output of the PID controller in the position loop. The robot's angular velocity is obtained from the output of the direction loop PID controller. This indicates the linear velocity of the robot's left wheel. This represents the linear velocity of the robot's right wheel, where W is the distance between the wheels, ultimately generating four-wheel drive commands.

[0063] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0064] This invention designs a multi-robot collaborative target recognition and tracking control method in dynamic environments. The system uses an improved YOLOv5s network as the core detection module to achieve real-time and accurate recognition of target personnel. Combining an optimized APF (Advanced Position Detection Function) and a dual-closed-loop PID control architecture, a closed loop from target recognition to motion control is achieved. After the system confirms the target position, each robot calculates its motion trajectory in real time based on the distributed APF. At the motion control level, a hierarchical PID design is adopted. The outer-loop pose controller converts the desired coordinates output by the potential field and the deviation from the current pose into speed commands, while the inner-loop speed controller achieves accurate tracking through encoder feedback, ultimately generating a differential drive signal. Attached Figure Description

[0065] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a schematic diagram of a multi-robot collaborative target recognition and tracking method in a dynamic environment according to an embodiment of the present invention;

[0067] Figure 2 This is a structural diagram of the CBRM module according to an embodiment of the present invention;

[0068] Figure 3 This is a structural diagram of the MobileNetV3 bottleneck block in an embodiment of the present invention, wherein (a) is a schematic diagram with stride=1; and (b) is a schematic diagram with stride=2.

[0069] Figure 4 This is a schematic diagram of the improved backbone network according to an embodiment of the present invention;

[0070] Figure 5 This is a diagram of the CA-YOLOv5s backbone network structure according to an embodiment of the present invention;

[0071] Figure 6 The diagrams show a comparison of different feature fusion structures in embodiments of the present invention, wherein (a) is a schematic diagram of the PANet structure; and (b) is a schematic diagram of BiFPN.

[0072] Figure 7 This is a schematic diagram of CIOU according to an embodiment of the present invention;

[0073] Figure 8 This is a schematic diagram of EIOU in an embodiment of the present invention;

[0074] Figure 9 This is a schematic diagram of the center point of a sub-block in an embodiment of the present invention;

[0075] Figure 10 This is a simplified search domain diagram according to an embodiment of the present invention;

[0076] Figure 11 This is a schematic diagram of the K-Means tree algorithm with priority search in an embodiment of the present invention;

[0077] Figure 12 This is a flowchart of the PROSAC algorithm according to an embodiment of the present invention;

[0078] Figure 13 This is a schematic diagram of the re-identification mechanism workflow according to an embodiment of the present invention;

[0079] Figure 14 This is a schematic diagram illustrating the local minimum value of APF in an embodiment of the present invention;

[0080] Figure 15 This is a schematic diagram of the APF after adding random perturbations in an embodiment of the present invention;

[0081] Figure 16 This is a schematic diagram illustrating the unreachable location relationship of the target in an embodiment of the present invention;

[0082] Figure 17 This is a schematic diagram illustrating the unreachable APF target situation according to an embodiment of the present invention;

[0083] Figure 18 This is a schematic diagram of the improved repulsive potential field (APF) according to an embodiment of the present invention.

[0084] Figure 19 This is a schematic diagram of the dual closed-loop PID controller according to an embodiment of the present invention;

[0085] Figure 20 This is a top view of the experimental environment in an embodiment of the present invention;

[0086] Figure 21 This is a schematic diagram of the preset trajectory of the target pedestrian in an embodiment of the present invention;

[0087] Figure 22This is a visualization diagram of the Rviz recognition effect in an embodiment of the present invention;

[0088] Figure 23 This is a schematic diagram of the multi-robot tracking control trajectory according to an embodiment of the present invention. Detailed Implementation

[0089] 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.

[0090] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0091] Example 1

[0092] The flowchart of the present invention is as follows Figure 1 As shown. This invention discloses a multi-robot collaborative target recognition and tracking method in a dynamic environment. It uses a YOLOv5s network to detect targets and personnel, achieving real-time and accurate recognition through its improved network structure. The tracking strategy is maintained when the confidence level is above a threshold; if the confidence level is below the threshold and a valid detection box exists, a re-identification module is activated, using a TCN network combined with temporal features and motion patterns for identity verification. The threshold is typically set empirically, usually between 0.7 and 0.8. For severely occluded or rapidly moving targets, the system activates a multi-view collaborative mechanism, utilizing visual information from multiple robot platforms to improve the robustness of recognition. When a change in target features is detected, the system updates model parameters in real time to adapt to the new features. Furthermore, the system employs an optimized APF and PID control strategy to establish a closed-loop system from target recognition to motion control. After confirming the target position, the robot calculates its trajectory based on a distributed APF; the gravitational component guides the robot to converge towards the target, while the repulsive component avoids obstacles. In terms of control, a hierarchical PID architecture is adopted. The outer-loop pose controller converts the desired coordinates into velocity commands, and the inner-loop velocity controller tracks the target through encoder feedback, generating differential drive signals. The entire architecture shares robot pose information through a ROS network to dynamically respond to changes in the positions of multiple robots. Finally, experiments demonstrated the effectiveness of the invention.

[0093] In this embodiment, the target personnel are identified in real time:

[0094] This invention, based on the YOLOv5s model, enables real-time target person identification. Current YOLOv5s neck network employs an FPN + PAN structure. On one hand, the predominantly unidirectional fusion approach (top-down FPN combined with bottom-up PAN) leads to feature loss during transmission, resulting in insufficient fusion of multi-scale features, especially small target features, impacting small target detection accuracy and leading to a high false negative rate for small obstacles and targets in complex scenes. On the other hand, upsampling high-order feature maps using nearest-neighbor interpolation only considers adjacent sub-pixel spaces, resulting in a small receptive range. This is detrimental to feature extraction from distant targets in noisy environments and can cause color blockages, further reducing small target detection accuracy. Furthermore, this upsampling algorithm does not utilize semantic information from the feature map, resulting in a small receptive field and inability to effectively aggregate surrounding information. The lack of an integrated attention mechanism prevents targeted enhancement of key features, leading to insufficient extraction and utilization of useful features in complex environments, thus affecting detection performance.

[0095] To address the resource constraints faced by YOLOv5s models when deployed on embedded devices, a lightweight improvement scheme is proposed. First, the backbone network is reconstructed using MobileNetV3, significantly reducing computational complexity. Second, a novel integration of the CA attention mechanism and BiFPN structure in the neck network significantly improves multi-scale feature fusion efficiency. Finally, the detection head performance is optimized by introducing the EIOU loss function. The specific implementation steps are as follows:

[0096] Step 1: Improvements to the MobileNetV3 backbone network. The specific implementation process is as follows: First, replace the original network's Focus module and the first Conv and C3 modules with a CBRM structure. This structure includes a convolutional layer (Conv(64,3,2)) with 64 input channels, a kernel size of 3, and a stride of 2, and a max pooling layer (MaxPool) with a kernel size of 2, a stride of 3, and a padding value of 1. Figure 2 As shown.

[0097] Then, an improved MobileNetV3 bottleneck structure is used for network reconstruction, employing a cascaded design: first, 1×1 pointwise convolutions and h-Swish activation are applied, followed by deepwise convolutions with a stride of 2 for downsampling; then, an SE attention mechanism is introduced to enhance key features; finally, pointwise convolutions are used for dimensionality reduction. Figure 3 As shown in (a), the feature extraction module retains residual connections, the left branch maintains a depthwise convolution with stride=1, and finally performs feature fusion with the shortcut branch. In particular, residual connections are canceled during downsampling to maintain feature dimension matching, and all modules use the h-Swish activation function to balance computational efficiency and non-linear expressive power. Figure 3 (b) is a downsampling module structure in a lightweight convolutional neural network. It forces downsampling by stride=2 and removes the ADD residual connection. It reduces computational power through depthwise separable convolution, focuses on key points through SE attention, and preserves gradients through h-Swish, enabling the model to efficiently process image features on mobile devices.

[0098] By reconstructing the backbone network using the MobileNetV3 bottleneck structure, the model performance was significantly optimized. On the one hand, the application of depthwise separable convolutions and inverse residual structures effectively reduced parameter redundancy and computational cost, thus lowering the model complexity. On the other hand, the introduction of the SE attention mechanism enhanced the network's ability to extract key features and improved the model's representation efficiency. The SPPF module was used for multi-scale feature extraction to enhance the model's ability to detect targets of different sizes. Figure 4 The improved lightweight backbone network structure is demonstrated, which achieves a significant improvement in computational efficiency without sacrificing detection accuracy.

[0099] Step 2: Improvement of the neck network based on CA and BiFPN. By embedding the CA attention mechanism module into the backbone structure of YOLOv5s, the improved network model YOLOv5s-CA is obtained, as shown below. Figure 5 As shown.

[0100] The original PANet is replaced by a BiFPN structure. The structure is as follows: Figure 6 As shown. Figure 6 (a) Figure 6 (b) BiFPN, as a weighted bidirectional feature pyramid network, removes isolated nodes at higher levels to simplify the network and introduces weighted connections between input and output nodes at the same level, as shown by the arrows in the figure, to achieve learnable feature importance allocation and significantly improve feature fusion efficiency.

[0101] Taking the second-layer BiFPN node as an example, the calculation process of its feature fusion is shown in Equations (1) and (2).

[0102] (1)

[0103] (2)

[0104] in, Represents intermediate fusion features, and Conv represents the convolution operation. , Represents the trainable parameters obtained through learning, and Resize represents the size adjustment operation. Represents the input feature map, Represents a deeper input feature map. Represents the final output features. , , Represents trainable parameters, Represents cross-layer characteristics, This represents the minimum value, preventing the denominator from being zero and causing calculation crashes.

[0105] Step 3: Improvement of the head network based on the loss function, using the CIOU loss function, such as... Figure 7 As shown.

[0106] CIOU is defined as:

[0107] (3)

[0108] The CIOU loss function is:

[0109] (4)

[0110] in, These are weighting coefficients used to balance the influence of the aspect ratio. It measures the difference in aspect ratio between the predicted bounding box and the actual bounding box.

[0111] The EIOU loss function is a key improvement over CIOU, enhancing positioning accuracy by introducing a more refined aspect ratio optimization mechanism. For example... Figure 8 As shown.

[0112] EIOU is defined as:

[0113] (5)

[0114] The EIOU loss function is:

[0115] (6)

[0116] Among them, b and These are the center coordinates of the anchor box and the real bounding box, respectively. This represents the Euclidean distance between the two, where c is the diagonal distance of the minimum bounding rectangle. and The minimum bounding rectangle width and height. , Define the width and height of the marker box; and It refers to the width and height of the anchor frame.

[0117] To address the challenges of target detection in complex scenarios, the EIOU loss function is used instead of the original CIOU loss function. Its excellent aspect ratio adaptability can effectively improve the detection accuracy of targets of random sizes.

[0118] In this embodiment, multi-angle image collaborative image stitching:

[0119] This invention constructs a global panoramic perception view through multi-view image stitching and distributed target fusion technology, effectively eliminating blind spots in the field of view of a single robot and expanding the monitoring coverage. Simultaneously, it utilizes the complementary characteristics of multiple robot perspectives to solve the tracking interruption problem caused by target occlusion, improving the system's robustness and tracking continuity in dynamic environments. Specifically, it is implemented through the following steps:

[0120] Step 1: In the image preprocessing stage, this invention introduces an advanced color transfer algorithm to achieve color consistency correction between multi-view images by establishing a color mapping relationship between the source image and the target image. First, the mean and variance of each component in the color space of the two images are calculated, and then the following linear transformation relationship is established, as shown in equation (7).

[0121] (7)

[0122] In the image conversion process shown in equation (7), both the source image and the target image consist of three color channels. Specifically, the three-channel raw data of the source image is represented as ( , , ), whose corresponding mean and standard deviation are ( ) , , )and( , , Similarly, the output data of the target image is (L, A, B), and its statistical characteristic parameters include the mean ( , , ) and standard deviation ( , , ).

[0123] Step 2: One of the bottlenecks in the computational efficiency of the ORB algorithm stems from the fact that its FAST feature point detection mechanism requires traversing and searching all pixels in the entire image. To address the efficiency optimization issue of the FAST feature detection algorithm, an improved method based on image region filtering is proposed. Since corner points detected by the FAST algorithm are usually distributed in areas with significant pixel value changes, a quadtree decomposition technique is introduced. By analyzing the gray-level distribution characteristics of the image, regions with gradual gray-level changes in both the reference image and the image to be registered are automatically filtered out. This region filtering mechanism effectively narrows the search range for feature point detection, performing FAST detection only in texture-rich regions, thereby significantly reducing the time consumption in the feature extraction stage.

[0124] The original image region S is recursively divided into four equal parts, and the center point coordinates of each sub-block are accurately located in each division process, such as... Figure 9 As shown.

[0125] Determine the quadtree partitioning depth. To prevent the quadtree from being partitioned too deeply, which would introduce additional time consumption, it is necessary to determine the depth of the quadtree partitioning.

[0126] Considering that oFAST feature detection requires calculations based on a Bresenham circle with a radius of 3 pixels, the system sets the minimum processing unit area to 10 square pixels. When the area of ​​an image sub-region shrinks to this threshold (i.e., 10 square pixels), the decomposition process of the current branch will be automatically terminated.

[0127] For the target area to be retained First, determine the center point p1 of the region. Then, select two symmetrical points (p2-p5) in both the vertical and horizontal directions. The distance between these sampling points and the center point is half the width or height of the region (experiments have verified that this ratio has the best effect). Based on these five sampling points, calculate the region feature value using equation (8), where... This represents the pixel value of the j-th sampling point in the i-th sub-region. The feature value represents the i-th sub-region.

[0128] (8)

[0129] In the region selection stage, Manhattan distance is used as a similarity metric. As shown in equation (9), where D i This represents the Manhattan distance between pixels within the i-th sub-region. A threshold is set... (The value of z is selected based on the specific application scenario to distinguish between areas with uniform and uneven pixel distribution, thereby achieving the effect of region filtering.) Perform region filtering: When If the pixel distribution in a region is uniform, it is discarded; otherwise, it is retained and further decomposed.

[0130] (9)

[0131] An iterative optimization region selection strategy is adopted, and the above operation process is repeatedly executed on the retained candidate regions.

[0132] Figure 10 The diagram illustrates the search domain filtering process. This process employs a quadtree decomposition strategy, where a "T" marker indicates a region with significant pixel grayscale changes, requiring further decomposition; an "F" marker indicates a region with uniform grayscale distribution, which is then eliminated. The algorithm execution process is as follows: The original image S is used as the root node of the quadtree for initial partitioning. Grayscale changes are evaluated in each sub-region, and regions meeting the conditions are recursively decomposed. The final output is shown below. Figure 10The optimized search area is shown as d in the figure.

[0133] Step 3: With a fixed number of neighboring points N, the maximum threshold that ensures pixel P can still be identified as a corner point is defined as the corner response value of that point. For each candidate pixel P, perform the following operations: calculate the absolute grayscale difference between P and its 8 consecutive adjacent pixels on the circumference, and obtain the minimum value d; consider the first and last adjacent extended pixels, and calculate their absolute grayscale difference with P respectively. and The corner response value of the current iteration is determined by the comparison operation shown in equation (10). .

[0134] (10)

[0135] After a complete traversal and calculation of all 16 sampling points on the circumference, the corner response value of pixel P was finally determined as follows:

[0136] (11)

[0137] After obtaining the corner response value of each pixel, the response value of each candidate point is compared within its 3×3 neighborhood, and only those points with the largest response value in their 8-neighborhood are retained as the final corner points.

[0138] Step 4: To enhance the rotation invariance of the FAST algorithm and facilitate subsequent feature angle calculation, the Harris response value is introduced as an auxiliary evaluation metric. First, for each corner point detected by FAST, the Sobel operator is used to calculate the gradient values ​​of all pixels in the x and y directions within its neighborhood, denoted as... and Then, the gradient combination quantities a, b, and c are calculated according to equations (13)-(15) respectively. Finally, the response intensity of each corner point is obtained through the Harris response function shown in equation (12), where k is an empirical coefficient and s is a scale parameter. This fusion strategy retains the efficiency of the FAST algorithm and enhances the rotation robustness of the algorithm through the feature analysis of the Harris matrix.

[0139] (12)

[0140] (13)

[0141] (14)

[0142] (15)

[0143] The corner features detected by the FAST algorithm are characterized by the intensity of the offset vector. The ORB algorithm uses the intensity centroid method to determine the corner direction. First, it calculates the spatial moments of the image block to establish the direction reference. This moment reflects the gray-level distribution characteristics of the local region, as shown in Equation (16).

[0144] (16)

[0145] In the formula Represents pixels The gray value at the location, p, q represent the order of the moment, and the centroid of the image patch can be determined by calculating the first moment. Coordinates:

[0146] (17)

[0147] In the formula Represents the zeroth-order space moment. Represents the first spatial moment in the x-direction. The first-order spatial moment in the y-direction is represented by a corner point, and a local neighborhood is defined with the corner point as the center. The principal direction of the feature point is directly determined by calculating the spatial distribution characteristics of this region. :

[0148] (18)

[0149] The FAST algorithm cannot effectively match the limitations of feature points when processing scaled images. This invention introduces the concept of a scale pyramid. First, a scaling factor is set. The value is typically 1.2, and the number of pyramid layers, L, is usually chosen to be 8. Based on these parameters, the original image is progressively scaled down to create a series of scaled image sets. As shown in equation (19):

[0150] (19)

[0151] in, The scale of the original image. The range of values ​​is .

[0152] A multi-scale strategy is adopted to extract feature points. After detecting oFAST features at each scale level through a pyramid model, the features are fused to ensure scale invariance.

[0153] Subsequently, the BRIEF algorithm is used to encode the feature points into compact binary descriptors. This binary representation method significantly improves the computational efficiency of feature matching while maintaining strong discriminative ability.

[0154] Step 5: After obtaining the improved ORB feature points, FLANN is used for fast nearest neighbor matching, and the PROSAC algorithm is used to remove mismatched points.

[0155] The FLANN algorithm library handles large-scale data matching problems using efficient approximate nearest neighbor search techniques. Addressing the memory constraints of embedded platforms, a feature matching optimization algorithm based on a priority search K-Means tree is proposed. Figure 11 As shown, the implementation steps are as follows:

[0156] Classic PROSAC requires data points to be sorted by similarity evaluation first, and the accuracy and rationality of the sorting directly affect subsequent sampling and model estimation. Classic PROSAC must handle issues such as the definition and growth of the hypothesis generation set; it cannot rely excessively on pre-sorting using evaluation functions, nor can it treat all data points equally. In practice, properly controlling the hypothesis generation set is difficult and can easily lead to model estimation bias or affect computational efficiency. When classic PROSAC is integrated with novel feature extraction and matching algorithms (such as deep feature matching), its adaptability and synergy are insufficient, making it difficult to fully leverage the combined advantages.

[0157] The flowchart of the improved PROSAC algorithm proposed in this invention is as follows: Figure 12 As shown, the specific implementation steps are as follows:

[0158] Step 1: Convert the feature point pair set obtained from FLANN coarse matching According to the matching quality function Sort in descending order to generate an ordered set of feature point pairs. .

[0159] (20)

[0160] in, This represents the matching quality score for the i-th feature point pair. This represents a pair of feature points.

[0161] Step 2: From The first n high-quality matching points are selected sequentially to form a set N. The quality standard of set N is defined as shown in equation (21):

[0162] ;(twenty one)

[0163] Step 3: Divide the feature point pairs into groups of 4 points each, sorting them by quality in descending order. Calculate the sum of the quality of each group and sort them. Select the 4 groups of matching points with the highest quality to calculate the homography matrix H.

[0164] ;(twenty two)

[0165] in and These are the coordinates of the matching point.

[0166] Step 4: Calculate the projection error of the remaining points using the homography matrix H.

[0167] The matching quality is evaluated by calculating the geometric error between the feature point and its projection point. If the error is less than the projection error threshold... If the point is an interior point, then it is considered an interior point; otherwise, it is considered an exterior point. Assume a point... The projection point is Then the projection error The calculation formula is shown in equation (23).

[0168] ;(twenty three)

[0169] Step 5: Projection Error Threshold Defined as:

[0170] ;(twenty four)

[0171] Where is the height of the H image, It is the width of the image. It is a constant, usually set to 0.1.

[0172] Step 6: Compare the number of interior points with the preset threshold (i.e., the threshold defined by the above formula (24)). If the condition is met, return the optimal model parameters; otherwise, continue iterating until the maximum number of iterations is reached.

[0173] Step 7: Homography describes the perspective mapping relationship of the same plane under different viewpoints. When a camera captures a planar scene from different angles, there is a geometric transformation between the images, which can be represented as a homography matrix H. Assume the image... Feature points in With images Features For matching point pairs, H is the homography transformation matrix, and the relationship between the two feature points is as follows:

[0174] (25)

[0175] In the above formula, H is a 3×3 matrix, for each pair of matching points... Substituting into the above equation, we get the following equation:

[0176] (26)

[0177] If is a parameter of the homography matrix, then further transformation yields the following formula:

[0178] (27)

[0179] Write the above expression as The system of equations takes the form of the following:

[0180] (28)

[0181] Solving for the homography matrix requires determining eight independent parameters. Since two equations can be established for each pair of matching points, theoretically, a unique solution can be found with four sets of non-collinear matching points. After eliminating redundant parameters through constraints, the optimal homography matrix is ​​calculated using the least squares method.

[0182] Step 8: The fade-in / fade-out blending algorithm achieves a smooth transition by dynamically adjusting the blending weights of pixels in overlapping areas. This algorithm calculates the blending ratio based on the distance between pixels and the image boundary; pixels closer to the boundary are assigned higher weights, resulting in a natural gradient effect in the stitched area. The specific expression is as follows:

[0183] (29)

[0184] In equation (29), This represents the pixel value at coordinates (x, y) in the fused output image. This represents the pixel value at coordinates (x, y) in the left-hand image. This represents the pixel value at coordinates (x, y) in the image on the right. This represents the fusion weight of the left image at coordinates (x, y). This represents the fusion weight of the image on the right at coordinates (x, y). and These represent the non-overlapping areas on the left and right sides, respectively. This represents the overlapping region. The pixel values ​​of the output image consist of three parts: the non-overlapping regions directly use the pixel values ​​of the original image, while the overlapping regions have their blending weights calculated based on pixel positions. The weight coefficients are determined by the normalized distance from the pixel to the boundary. The weight of the left image decreases as the distance to the right boundary increases, and the weight of the right image decreases as the distance to the left boundary increases; the sum of the two is always 1.

[0185] In this embodiment, target re-identification:

[0186] To address the target occlusion problem during the following process, a temporal convolutional network (TCN) is introduced on the basis of the existing person recognition network. Compared with recurrent neural network structures such as LSTM and GRU, the temporal convolutional network shows significant advantages in temporal modeling.

[0187] The construction process of the TCN-based target recapture scheme is as follows: Figure 13As shown, this mechanism integrates event trigger judgment and overlap rate calculation functions. When the system detects a target loss event, it automatically activates the re-acquisition process, achieving target relocation by analyzing historical motion trajectories and real-time appearance features.

[0188] This invention employs a TCN (Transient Channel Network) to construct a target re-identification mechanism, fully leveraging its temporal memory characteristics to achieve dual verification by fusing target motion trajectory and appearance features. Specifically, it is implemented through the following steps:

[0189] Step 1: Predict the output, network input It consists of the positional changes of the target bounding box at adjacent time steps, specifically represented as the geometric difference between the detection box at the current time and the previous time step. This temporal difference feature effectively captures the movement trajectory change pattern of the target.

[0190] (30)

[0191] The network input includes the target's horizontal and vertical positional offsets in the image coordinate system, as well as the changes in the length and width of the detection box. Through layer-by-layer feature extraction via temporal convolutional layers, the network ultimately outputs the predicted position and scale changes of the bounding box for future multi-step distances, forming a complete motion trajectory prediction result.

[0192] in, Let be the Y-axis coordinate at time t. Let be the X-axis coordinate at time t. Let t be the measurement frame width. Let t be the height of the detection frame. and They are respectively The changes of the target along the X and Y axes in the image coordinate system within a given time period. and They are respectively The changes in the width and height of the detection box within a given time step are extracted layer by layer by the TCN. The network ultimately outputs the predicted position and scale change of the bounding box for a certain future step, forming a complete motion trajectory prediction result, denoted as... :

[0193] (31)

[0194] in, The step size for network prediction. for The predicted trajectory at any given moment;

[0195] Step 2: Event triggering mechanism. Under normal tracking conditions, the system ignores the prediction output of the temporal network, maintains the basic classification process, and maintains a fixed-length... Timing sliding window When a target loss event is detected, the re-identification verification mechanism is automatically activated, and the motion trajectory prediction result is cross-validated with the appearance classification output.

[0196] (32)

[0197] in, for The predicted trajectory at any given moment;

[0198] Step 3: Overlap rate calculation; a credibility score index is introduced in the re-identification verification process. and The calculation formula is as follows:

[0199] (33)

[0200] (34)

[0201] By comparing the bounding boxes output by the detection network With re-identification prediction boxes The degree of overlap is considered in conjunction with a preset threshold for decision-making. When the overall credibility assessment result exceeds the set threshold, the target re-identification is deemed successful, and the predicted bounding box compensates for the missing detection area; otherwise, it is considered a matching failure. This threshold is 1.7, obtained through multiple experiments.

[0202] (35)

[0203] In this embodiment, multi-robot obstacle avoidance with potential field optimization:

[0204] This invention introduces a random perturbation mechanism, applying random forces to help the robot escape local potential energy traps. This invention introduces random forces. This provides additional power to help the robot escape when it becomes trapped. The expression for this stochastic force is:

[0205] (36)

[0206] in: It is the intensity coefficient of random force, used to control the magnitude of random disturbances; It is a random vector whose direction is in Uniformly distributed within a range, with sizes ranging from Randomly generated within the range.

[0207] After introducing random forces, the total resultant force acting on the robot can be expressed as the vector sum of attractive force, repulsive force, and random force:

[0208] (37)

[0209] in, Represents gravity. Indicates repulsive force;

[0210] By introducing random perturbations, the APF algorithm can apply random forces to help the robot escape when it gets trapped in a local minimum. For example... Figure 14 and Figure 15 As shown, the improved algorithm effectively solves the local minima problem.

[0211] On the other hand, when the obstacle's influence area includes the target point, as the robot approaches the obstacle, the repulsive force increases while the attractive force decreases, making the target unreachable. The positional relationship of the three is as follows: Figure 16 As shown.

[0212] When the robot, the target point, and the obstacle are collinear, and the target point is centered, the robot will stop at a certain equilibrium point where the net force is zero. Therefore, the repulsive potential field function is improved. As shown in equation (38), the repulsive potential energy at the target point is 0.

[0213] (38)

[0214] in: Indicates the order, ; This indicates that the target person is located in a repulsive potential field; It is the repulsive force gain coefficient; It's the robot's location; It indicates the location of the obstacle; It is the distance between the robot and the obstacle; It is the furthest distance affected by the repulsive potential field of the obstacle; Indicates the target location; Indicates the distance between the robot and the target;

[0215] The repulsive force acting on the robot after the improved repulsive potential field function is:

[0216] (39)

[0217] in, This represents the total repulsive force vector acting on the robot. The gradient represents the repulsive potential field. Represents the direct repulsive force of the obstacle. This represents the reverse repulsive force associated with the corresponding target point.

[0218] in , The calculation is as follows:

[0219] (40)

[0220] In the formula: The direction is that the obstacle is pointing towards the robot.

[0221] (41)

[0222] In the formula: The direction is that the robot is pointing at the target person.

[0223] like Figure 17 and Figure 18 As shown, the optimized algorithm ensures that the robot can successfully reach the target point.

[0224] In this embodiment, a dual-closed-loop PID controller is used:

[0225] This invention employs a parallel dual-closed-loop PID architecture, using two independent control channels to achieve precise adjustment of position and orientation angle. The schematic diagram of the dual-closed-loop PID is shown below. Figure 19 As shown.

[0226] In the position control channel, the outer loop PID controller follows the desired distance. Distance from actual distance deviation Generate target linear velocity The inner-loop PID controller is responsible for... To achieve precise tracking; in the direction control channel, the outer loop PID controller operates based on the desired azimuth angle. Compared with actual azimuth deviation Calculate the target angular velocity And implemented by inner-loop PID. The rapid response is achieved. The outputs of the two control channels are synthesized using a differential kinematics model. , ,in, The linear velocity of the robot is obtained from the output of the PID controller in the position loop. The robot's angular velocity is obtained from the output of the direction loop PID controller. This indicates the linear velocity of the robot's left wheel. This represents the linear velocity of the robot's right wheel, where W is the distance between the wheels, ultimately generating four-wheel drive commands.

[0227] In this embodiment, experimental verification and analysis are performed:

[0228] (1) To verify the robustness of the target tracking algorithm proposed in this invention, a comprehensive evaluation was conducted on a self-made data stream, which covers a variety of complex scenarios, including challenging conditions such as changes in illumination, scale changes, occlusion, target leaving the field of view, and background interference.

[0229] Experimental results show that the proposed method exhibits excellent tracking performance in various scenarios: it effectively addresses issues such as complex backgrounds and interference from similar pedestrians in bright environments; it handles the challenge of blurred target features under low-light conditions; and it maintains stable tracking capabilities even in extreme situations such as the target temporarily leaving the field of view of a single camera or encountering pedestrian occlusion. Especially in scenarios with drastic changes in lighting conditions and severe occlusion, the method can still accurately detect and continuously track the target, demonstrating strong environmental adaptability.

[0230] Experimental results show that the pedestrian following method proposed in this invention performs well in complex environments.

[0231] (2) ROS-Gazebo co-simulation experiment

[0232] This invention constructs an experimental environment based on the ROS Melodic and Gazebo 9.0.0 simulation platform, selecting the Turtlebot3 burger as the core experimental platform. This robot features a compact design with overall dimensions of 138mm x 178mm x 192mm (L x W x H), exhibiting good mobility. Although the Turtlebot3 burger employs a two-wheel differential drive structure, its kinematic model is equivalent to that of a four-wheel differential drive robot, both achieving steering control through the speed difference between the left and right wheels. Therefore, it can serve as an effective alternative platform for algorithm verification to four-wheel differential robots. Regarding sensor configuration, the burger robot is equipped with a Kinect RGB-D camera and an LDS-01 LiDAR sensor. The Kinect camera handles target recognition and feature extraction, while the LDS-01 LiDAR is used for environmental perception and obstacle detection. Their collaborative operation enables accurate localization and tracking of dynamic targets. The experiment involved deploying three Turtlebot3 burger robots in Gazebo, utilizing ROS's topic communication mechanism to achieve information interaction between the robots.

[0233] A 20m x 20m square simulation scene was constructed, with the area spatially calibrated using a 1m x 1m grid. The environment included two types of obstacles: static obstacles consisting of cross-shaped inner walls, and other pedestrians as dynamic obstacles. To simulate a realistic tracking scenario, the target person followed... Figure 20 , Figure 21 Move along the preset trajectory shown.

[0234] The system's recognition results are monitored in real time using the Rviz visualization tool, and the visualization effect is as follows: Figure 22 As shown.

[0235] The trajectories of the three robots are as follows Figure 23As shown, the results demonstrate that, guided by the tracking control strategy of this invention, each robot effectively avoided obstacles and achieved stable tracking of the target.

[0236] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for multi-robot cooperative target recognition and tracking in a dynamic environment, characterized in that, The method includes: Real-time identification of target personnel; If the confidence level of the target person is lower than the threshold and there is a valid detection box, the re-identification module is activated, and the TCN network is used to combine temporal features and motion patterns to confirm the identity. If the target is lost, a multi-view collaboration mechanism is activated. Through multi-view image stitching and distributed target fusion technology, a global panoramic perception view is constructed. At the same time, the complementary characteristics of multiple robot perspectives are utilized to solve the tracking interruption problem caused by target occlusion. If the confidence level of the target person is higher than the threshold, the target position is acquired in real time, and an optimized APF and PID control strategy is used to establish a closed-loop system from target recognition to motion control to maintain the tracking strategy. If the confidence level of the target person is below the threshold and a valid detection box exists, the re-identification module is activated. Methods for identity verification using a TCN network combined with temporal features and motion patterns include: Predicted output: Network input It consists of the positional changes of the target bounding box at adjacent time points, specifically represented by the current time point. Compared to the previous moment Geometric difference of the detection frame: ； The network input includes the target's horizontal and vertical positional offsets in the image coordinate system, as well as the changes in the length and width of the detection box. Through layer-by-layer feature extraction via temporal convolutional layers, the network ultimately outputs the predicted position and scale changes of the bounding box for future multi-step distances, forming a complete motion trajectory prediction result. in, for t Y-axis coordinate at time, for t X-axis coordinate at time, for t Detect frame width dimensions at all times. for t Continuously monitor the frame height dimension. and They are respectively The changes of the target along the X and Y axes in the image coordinate system within a given time period. and They are respectively The changes in the width and height of the bounding box within a given time step are analyzed through layer-by-layer feature extraction by the TCN. The final output is the predicted position of the bounding box at the preset step size, along with the scale change, forming a complete motion trajectory prediction result, denoted as... : ； in, The step size for network prediction. for The predicted trajectory at any given moment; Event triggering mechanism: Under normal tracking conditions, the system ignores the prediction output of the temporal network, maintains the basic classification process, and keeps a fixed length. Timing sliding window When a target loss event is detected, a re-identification verification mechanism is activated, cross-validating the motion trajectory prediction result with the appearance classification output; among which, ； in, for The predicted trajectory at any given moment; Overlap rate calculation: A credibility score index is introduced into the re-identification verification process. and The calculation formula is as follows: ； ； By comparing the bounding boxes output by the detection network With re-identification prediction boxes The degree of overlap is considered in conjunction with a preset threshold for decision-making. When the overall credibility assessment result exceeds the set threshold, the target re-identification is deemed successful, and the predicted bounding box compensates for the detected missing region; otherwise, it is considered a matching failure. ； If the confidence level of the target person is higher than the threshold, the target position is acquired in real time. An optimized APF and PID control strategy is used to establish a closed-loop system from target recognition to motion control. The methods for maintaining the tracking strategy include: After confirming the target location, the robot calculates its trajectory based on the distributed APF. The gravitational component guides the robot to converge toward the target, while the repulsive component avoids obstacles. A hierarchical PID architecture is adopted. The outer loop posture controller converts the desired coordinates into speed commands, and the inner loop speed controller tracks the target through encoder feedback and generates differential drive signals.

2. The method according to claim 1, characterized in that, Methods for real-time identification of target personnel include: Based on the YOLOv5s model, the task of real-time identification of target personnel is achieved; The YOLOv5s model uses MobileNetV3 to reconstruct the backbone network, integrates the CA attention mechanism and BiFPN structure in the neck network, and optimizes the detection head performance by introducing the EIOU loss function.

3. The method according to claim 1, characterized in that, If the target is lost, a multi-view collaborative mechanism is activated. This mechanism uses multi-view image stitching and distributed target fusion technology to construct a global panoramic perception view. Simultaneously, leveraging the complementary characteristics of multiple robot perspectives, methods to address tracking interruptions caused by target occlusion include: A color transfer algorithm is introduced in the image preprocessing stage to establish a color mapping relationship between the source image and the target image; The original image region S is recursively divided into four equal parts, and the center point coordinates of each sub-block are located in each division process; Under the condition of a fixed number of neighborhood points N, the maximum threshold that allows a pixel point P to still be identified as a corner point is defined as the corner response value of that point. The Harris response value is introduced as an auxiliary evaluation index. The FAST algorithm is used to characterize the corner intensity through the offset vector. The ORB algorithm is used to determine the corner direction through the intensity centroid. The concept of scale pyramid is introduced and implemented by generating images at multiple scale levels. After obtaining the improved ORB feature points, FLANN is used for nearest neighbor matching, and the PROSAC algorithm is used to remove mismatched points.

4. The method according to claim 3, characterized in that, After obtaining the improved ORB feature points, the method of using FLANN for nearest neighbor matching and combining it with the PROSAC algorithm to remove mismatched points includes: The set of feature point pairs obtained from FLANN coarse matching According to the matching quality function Sort in descending order to generate an ordered set of feature point pairs. ; from Select the first one in order n A set of high-quality matching points N Define a set N Quality standards; Feature point pairs are grouped into groups of four based on descending quality. The sum of the quality of each group is calculated and sorted. The four groups with the highest quality are selected to calculate the homography matrix. H ; Using homography matrix H Calculate the projection error of the remaining points; The matching quality is evaluated by calculating the geometric error between the feature point and the corresponding projection point. If the error is less than the threshold, the feature point is regarded as an interior point; otherwise, it is regarded as an exterior point. Compare the number of inliers with a preset threshold. If the condition is met, return the optimal model parameters; otherwise, continue iterating until the maximum number of iterations is reached.

5. The method according to claim 1, characterized in that, After confirming the target location, the robot calculates its trajectory based on a distributed APF (Automatic Persistent Force). The gravitational component guides the robot to converge toward the target, while the repulsive component helps it avoid obstacles. Introducing random forces This provides additional power when the robot is stuck; the expression for this stochastic force is: ； in: It is the intensity coefficient of random force, used to control the magnitude of random disturbances; It is a random vector, with the direction in Uniformly distributed within a range, with sizes ranging from Randomly generated within the range; After introducing random forces, the total resultant force acting on the robot can be expressed as the vector sum of attractive force, repulsive force, and random force: ； in, Represents gravity. Indicates repulsive force; When the robot, the target point, and the obstacle are collinear, and the target point is centered, the robot will stop at a preset equilibrium point, where the net force is 0. The repulsive potential field function is then improved to make the repulsive potential energy at the target point zero. The repulsive potential field function is: ； in, Indicates the order, ; This indicates that the target person is located in a repulsive potential field; It is the repulsive force gain coefficient; It's the robot's position; It indicates the location of the obstacle; It is the distance between the robot and the obstacle; It is the furthest distance affected by the repulsive potential field of the obstacle; Indicates the target location; Indicates the distance between the robot and the target; The repulsive force acting on the robot after the improved repulsive potential field function is: ； in, This represents the total repulsive force vector acting on the robot. The gradient represents the repulsive potential field. Represents the direct repulsive force of the obstacle. This represents the reverse repulsive force associated with the corresponding target point.

6. The method according to claim 1, characterized in that, A hierarchical PID architecture is adopted, in which the outer-loop pose controller converts the desired coordinates into velocity commands, and the inner-loop velocity controller tracks the target through encoder feedback. The method for generating differential drive signals includes: In the position control channel, the outer loop PID controller follows the desired distance. Distance from actual distance deviation Generate target linear velocity The inner loop PID is responsible for... Track; In the directional control channel, the outer loop PID controller operates based on the desired azimuth angle. Compared with actual azimuth deviation Calculate the target angular velocity And implemented by inner-loop PID. Rapid response; The outputs of the two control channels are synthesized using a differential kinematics model: , ,in, The linear velocity of the robot is obtained from the output of the PID controller in the position loop. The robot's angular velocity is obtained from the output of the direction loop PID controller. This indicates the linear velocity of the robot's left wheel. This indicates the linear velocity of the robot's right wheel. W The distance between the wheels is used to generate the final four-wheel drive command.

Images