Multi-robot cooperative target identification and tracking method in dynamic environment

Abstract

The invention discloses a multi-robot cooperative target identification and tracking method in a dynamic environment. The method comprises the following steps: identifying a target person in real time; if the confidence coefficient of the target person is lower than a threshold value and an effective detection frame exists, a re-identification module is started, and identity confirmation is carried out by utilizing a TCN network in combination with time sequence characteristics and a motion mode; if the target is lost, a multi-view collaborative mechanism is started, a global panoramic perception view is constructed through a multi-view image splicing and distributed target fusion technology, and meanwhile, the problem of tracking interruption caused by target shielding is solved by utilizing the multi-robot view complementary characteristic; if the confidence coefficient of the target person is higher than a threshold value, the target position is obtained in real time, an optimized APF and PID control strategy is adopted to establish a closed-loop system from target recognition to motion control, and a tracking strategy is kept. According to the method, the problems of view shielding, target loss and dynamic interference in a complex environment are effectively solved, and high-precision and high-stability target tracking is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of multi-robot cooperative perception and intelligent decision-making, and particularly relates to a multi-robot cooperative target identification and tracking method in a dynamic environment. BACKGROUND

[0002] At present, the target following technology based on multi-sensor fusion has made significant progress. Researchers mainly use laser radar, visual camera and ultra-wideband (UWB) technology to realize target detection and positioning, and on this basis, develop corresponding robot navigation control algorithms. Among them, the point cloud data generated by laser radar is relatively sparse, which is difficult to capture the details of the target, such as the appearance, texture, etc. of the target, and in the target tracking task, it needs high computing resources to process the point cloud data captured by the laser radar; UWB usually needs the target to carry a tag (such as UWB beacon), which is limited by the equipment, and the positioning accuracy of UWB is affected by the environmental multipath effect (signal reflection), which may decrease in complex environment. However, the camera has the advantages of small size, low cost, and can capture high-resolution images, providing rich texture and color information, which is suitable for target identification and tracking tasks. Therefore, the research on autonomous tracking of designated personnel based on visual sensors is on the rise.

[0003] For the task of multi-robot cooperative target identification and tracking control, scholars use different methods to deal with it, such as the application of hdl_graph_slam system developed by Koide et al. in hospital navigation, but it is limited by the need to pre-map; Tasaki et al. perceive through the fusion of laser radar and IMU and verify the flexibility of the scheme. Pang et al. further integrate visual and laser data and use A* and TEB algorithm combination to complete target following, but this kind of method has the defect of insufficient real-time due to the need for complete path planning. Sun and Jun et al. adopted a dual-mode control strategy, dividing the following process into two independent modules of target tracking and obstacle avoidance, and through intelligent switching mechanism to ensure safe following. However, the motion discontinuity caused by speed regulation in this mode conversion process still needs to be further optimized.

[0004] The present application researches on the key technical problems of multi-robot cooperative target identification and tracking control in dynamic environment, and constructs a complete solution integrating lightweight detection, robust tracking and intelligent cooperative control. It has significant advantages in detection accuracy, tracking robustness and cooperative control efficiency. SUMMARY

[0005] To solve the problems in the prior art, the application provides a multi-robot cooperative target identification and tracking method in a dynamic environment, which effectively overcomes the problems of visual field obstruction, target loss and dynamic interference in a complex environment through multi-robot cooperative perception and intelligent decision-making, and realizes high-precision and high-stability target tracking, and can provide decision-making for multi-robot target identification and tracking in a dynamic environment.

[0006] To achieve the above object, the application provides the following scheme:

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

[0008] Real-time identification of target personnel;

[0009] If the confidence of the target personnel is lower than a threshold value and there is an effective detection frame, a re-identification module is started, and TCN network is used to combine time sequence features and motion patterns to confirm the identity;

[0010] If the target is lost, a multi-view cooperative mechanism is started, a global panoramic perception view is constructed through multi-view image splicing and distributed target fusion technology, and meanwhile, the multi-robot view complementary characteristics are used to solve the problem of tracking interruption caused by target obstruction;

[0011] If the confidence of the target personnel is higher than a threshold value, 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 identification to motion control, and the tracking strategy is maintained.

[0012] Preferably, the method for real-time identification of target personnel comprises:

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

[0014] The YOLOv5s model adopts a MobileNetV3 restructured backbone network, a CA attention mechanism and a BiFPN structure are fused in the neck network, and an EIOU loss function is introduced to optimize the performance of the detection head.

[0015] Preferably, if the confidence of the target personnel is lower than a threshold value and there is an effective detection frame, a re-identification module is started, and TCN network is used to combine time sequence features and motion patterns to confirm the identity.

[0016] Prediction output: network input The position change amount of the target boundary frame at adjacent moments is composed, and is specifically represented as the geometric difference amount of the detection frame at the current moment and the previous moment:

[0017] ;

[0018] The network input includes the horizontal and vertical position offsets of the target in the image coordinate system, as well as the changes in the length and width of the detection box. Through the layer-by-layer feature extraction of the temporal convolution layer, the network finally outputs the predicted position value and scale change of the bounding box for multiple steps in the future, forming a complete motion trajectory prediction result.

[0019] in, is the Y-axis coordinate at time t, is the X-axis coordinate at time t, is the detection frame width at time t, is the height of the detection frame at time t, and They are The change of the target on the X-axis and Y-axis in the image coordinate system within a certain time. and They are The change in the width and height of the detection box within the moment is extracted layer by layer through TCN, and the predicted position value and scale change of the bounding box with the preset step size in the future are finally output to form a complete motion trajectory prediction result, which is recorded as :

[0020] ;

[0021] in, is the step size of network prediction, for Motion trajectory prediction results at each moment;

[0022] Event trigger mechanism: In the normal tracking state, the system ignores the prediction output of the time series network, keeps the basic classification process running, and maintains a fixed length Time series sliding window ,When a target loss event is detected, the re-identification verification mechanism is activated to cross-validate the motion trajectory prediction results with the appearance classification output;

[0023] ;

[0024] in, for Motion trajectory prediction results at each moment;

[0025] Overlap rate calculation: Introducing a credibility score indicator into the re-identification verification phase and , the calculation formula is as follows:

[0026] ;

[0027] ;

[0028] By comparing the bounding boxes output by the detection network and re-identification prediction box The overlap degree is combined with the preset threshold to make a decision. When the comprehensive credibility evaluation result exceeds the set threshold, the target re-identification is judged to be successful, and the predicted bounding box is used to compensate for the missing detection area. Otherwise, it is considered a match failure.

[0029] .

[0030] Preferably, if the target is lost, a multi-perspective collaborative mechanism is activated to construct a global panoramic perception view through multi-perspective image stitching and distributed target fusion technology. At the same time, the complementary characteristics of multiple robot perspectives are utilized to solve the tracking interruption problem caused by target occlusion. The following methods are used:

[0031] In the image preprocessing stage, a color migration algorithm is introduced to establish a color mapping relationship between the source image and the target image;

[0032] Recursively divide the original image area S into four equal parts, and locate the coordinates of the center point of each sub-block in each division process;

[0033] Under the condition of a fixed number of neighborhood points N, the maximum threshold that can keep the pixel point P still being judged as a corner point is defined as the corner point response value of the point;

[0034] Harris response value is introduced as an auxiliary evaluation index, the FAST algorithm is used to characterize the intensity of corner points through the offset vector, the ORB algorithm is used to determine the direction of corner points through the intensity centroid, and the concept of scale pyramid is introduced to achieve this 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 combined to eliminate mismatched points.

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

[0037] The feature point pair set obtained by FLANN rough matching , according to the matching quality function Sort in descending order to generate an ordered set of feature point pairs ;

[0038] from In the example, the first n high-quality matching points are selected in sequence to form a set N, and the quality standard of the set N is defined;

[0039] Divide the feature point pairs into groups of 4 points in descending order of quality, calculate the sum of the quality of each group and sort them, and select the 4 groups of matching points with the highest quality to calculate the homography matrix H;

[0040] Calculate the projection error of the remaining points using the homography H;

[0041] Calculate the geometric error between the feature points and the corresponding projected points to evaluate the matching quality, if the error is less than a threshold, the feature points are considered as inliers, otherwise as outliers;

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

[0043] Preferably, if the confidence of the target person is higher than a threshold, the target position is obtained 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 tracking strategy includes:

[0044] After confirming the target position, the robot calculates the motion trajectory based on distributed APF, the attractive component guides the robot to converge to the target, and the repulsive component avoids obstacles;

[0045] A hierarchical PID architecture is used, the outer loop position controller converts the desired coordinates into speed commands, and the inner loop speed controller tracks the target through encoder feedback to generate differential drive signals.

[0046] Preferably, after confirming the target position, the robot calculates the motion trajectory based on distributed APF, the attractive component guides the robot to converge to the target, and the repulsive component avoids obstacles. The method includes:

[0047] Introducing a random force When the robot is trapped, an additional force is provided, and the random force expression is:

[0048] ;

[0049] Where: is the strength coefficient of the random force, used to control the size of the random disturbance; is a random vector, the direction is uniformly distributed in , and the size is randomly generated in ;

[0050] After introducing the random force, the total force experienced by the robot is represented as the vector sum of the attractive force, repulsive force and random force:

[0051] ;

[0052] Where, represents the attractive force, represents the repulsive force;

[0053] When the robot, target point and obstacle are collinear and the target point is centered, the robot will be at a preset balance point, the resultant force is 0, and the repulsive potential field function is improved so that the repulsive potential energy at the target point is 0, wherein the repulsive potential field function is:

[0054] ;

[0055] wherein, represents the order, ; represents that the target personnel is located in the repulsive potential field; is a repulsive gain coefficient; is the robot position; is the obstacle position; is the distance between the robot and the obstacle; is the farthest influence distance of the obstacle in the repulsive potential field; represents the target position; represents the distance between the robot and the target;

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

[0057] ;

[0058] wherein, represents the total repulsive force vector received by the robot, represents the gradient of the repulsive potential field, represents the direct repulsive force of the obstacle, represents the corresponding reverse repulsive force associated with the target point.

[0059] Preferably, a layered PID architecture is adopted, the outer loop position controller converts the desired coordinates into speed instructions, and the inner loop speed controller tracks the target through encoder feedback to generate a differential drive signal. The method comprises:

[0060] In the position control channel, the outer loop PID generates the target linear speed according to the deviation between the desired following distance and the actual distance , and the inner loop PID is responsible for tracking ;

[0061] In the direction control channel, the outer loop PID calculates the target angular velocity by the deviation between the desired azimuth angle and the actual azimuth angle , and the inner loop PID realizes the fast response of ;

[0062] The outputs of the two control channels are synthesized through a differential kinematics model: , where, is the linear velocity of the robot, obtained from the position loop PID output, is the angular velocity of the robot, obtained from the direction loop PID output, represents the linear velocity of the left wheel of the robot, represents the linear velocity of the right wheel of the robot, and W is the wheel distance, and finally generates four-wheel drive instructions.

[0063] Compared with the prior art, the beneficial effects of the present application are:

[0064] The present application designs a multi-robot cooperative target recognition and tracking control method in a dynamic environment. The system uses an improved YOLOv5s network as the core detection module to realize real-time and accurate identification of target personnel. Combined with the optimized APF and double closed-loop PID control architecture, a closed loop from target recognition to motion control is realized. After the system confirms the target position, each robot calculates the motion trajectory in real time according to the distributed APF. The motion control layer adopts a hierarchical PID design. The outer loop pose controller converts the expected coordinates output by the potential field and the current pose deviation into a speed command. The inner loop speed controller realizes accurate tracking through encoder feedback, and finally generates a differential drive signal. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings described in the following are only some embodiments of the present application, and other drawings can also be obtained according to these drawings without creative labor for those skilled in the art.

[0066] Figure 1 It is a process flow diagram of a multi-robot cooperative target recognition and tracking method in a dynamic environment according to an embodiment of the present application.

[0067] Figure 2 It is a CBRM module structure diagram according to an embodiment of the present application.

[0068] Figure 3 It is a MobileNetV3 bottleneck block structure diagram according to an embodiment of the present application. (a) is a schematic diagram of stride=1; (b) is a schematic diagram of stride=2.

[0069] Figure 4 It is a schematic diagram of an improved backbone network according to an embodiment of the present application.

[0070] Figure 5 It is a CA-YOLOv5s backbone network structure diagram according to an embodiment of the present application.

[0071] Figure 6 For different feature fusion structure of the embodiment of the application, (a) is a schematic diagram of PANet structure; (b) is a schematic diagram of BiFPN;

[0072] Figure 7 For the CIOU schematic diagram of the embodiment of the application;

[0073] Figure 8 For the EIOU schematic diagram of the embodiment of the application;

[0074] Figure 9 For the sub-block center point schematic diagram of the embodiment of the application;

[0075] Figure 10 For the simplified search domain schematic diagram of the embodiment of the application;

[0076] Figure 11 For the priority search K-Means tree algorithm schematic diagram of the embodiment of the application;

[0077] Figure 12 For the PROSAC algorithm flowchart of the embodiment of the application;

[0078] Figure 13 For the re-identification mechanism workflow schematic diagram of the embodiment of the application;

[0079] Figure 14 For the APF local minimum situation schematic diagram of the embodiment of the application;

[0080] Figure 15 For the APF schematic diagram of the embodiment of the application after adding random disturbance;

[0081] Figure 16 For the target unreachable position relationship schematic diagram of the embodiment of the application;

[0082] Figure 17 For the APF target unreachable situation schematic diagram of the embodiment of the application;

[0083] Figure 18 For the APF schematic diagram of the embodiment of the application of improved repulsive potential field;

[0084] Figure 19 For the double closed loop PID principle diagram of the embodiment of the application;

[0085] Figure 20 For the experimental environment overhead view of the embodiment of the application;

[0086] Figure 21 For the target pedestrian preset trajectory schematic diagram of the embodiment of the application;

[0087] Figure 22A Rviz recognition effect visualization schematic diagram of an embodiment of the present application;

[0088] Figure 23 A multi-robot tracking control running track schematic diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0089] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0090] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0091] Embodiment one

[0092] The flowchart of the present application is shown in the figure. Figure 1 The present application discloses a multi-robot cooperative target recognition and tracking method in a dynamic environment. The YOLOv5s network is used to detect target personnel, the improved network structure is used to realize real-time accurate recognition, and the tracking strategy is maintained when the confidence is higher than the threshold. If the confidence is lower than the threshold and there is an effective detection frame, the re-identification module is started, the TCN network is used to combine the timing characteristics and motion mode to confirm the identity, and the threshold is generally set according to experience, and is usually taken as 0.7~0.8. For the target with serious occlusion or rapid movement, the system will start the multi-view cooperative mechanism, and the visual information of the multi-robot platform is used to improve the robustness of the recognition. When the target feature is detected, the system will update the model parameters in real time to adapt to the new features. In addition, the system uses the 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 the motion trajectory based on the distributed APF, the attractive component guides the robot to converge to the target, and the repulsive component avoids obstacles. In terms of control, a hierarchical PID architecture is used, the outer loop position controller converts the expected coordinates into a speed command, and the inner loop speed controller tracks the target through encoder feedback to generate a differential drive signal. The entire architecture shares the robot pose information through the ROS network to dynamically respond to the position changes of the multi-machine. Finally, the effectiveness of the present application is demonstrated through experiments.

[0093] In the present embodiment, the target personnel is recognized in real time:

[0094] The application is based on YOLOv5s model, realizes real-time identification of target personnel task. The current YOLOv5s neck network adopts FPN+PAN structure, on the one hand, the one-way fusion is mainly used (top-down FPN combined with bottom-up PAN), the feature is easy to lose in transmission, the multi-scale feature, especially the small target feature, is not fully fused, the small target detection precision is affected, and the miss detection rate of small obstacles and small targets in complex scenes is high. On the other hand, the nearest neighbor interpolation method is used for upsampling high-order feature maps, only the adjacent subpixel space is considered, the perception range is small, which is not conducive to the extraction of target features in long distance and high background noise, and the color block phenomenon will cause feature loss and reduce the small target detection precision; and the upsampling algorithm does not use semantic information of the feature map, the receptive field is small, and it cannot effectively aggregate the surrounding information. The attention mechanism is not integrated, the key features cannot be targetedly strengthened, the useful features cannot be extracted and utilized in the face of complex environment, and the detection effect is affected.

[0095] YOLOv5s model faces resource constraints when deployed on embedded devices, a set of systematic lightweight improvement scheme is proposed. First, the MobileNetV3 is used to reconstruct the backbone network, which greatly reduces the computational complexity; second, the CA attention mechanism and BiFPN structure are innovatively fused in the neck network, which significantly improves the multi-scale feature fusion efficiency; finally, the EIOU loss function is introduced to optimize the performance of the detection head. The following steps are used to realize it:

[0096] Step1: improvement of the backbone network based on MobileNetV3, the specific implementation process is as follows: first, replace the Focus module, the first Conv and C3 module of the original network with the CBRM structure, which contains a convolution layer (Conv (64, 3, 2)) with an input channel number of 64, a kernel size of 3 and a stride of 2, and a maximum pooling layer (MaxPool) with a kernel size of 2, a stride of 3 and a padding value of 1, as shown in Figure 2 .

[0097] Then, the improved MobileNetV3 bottleneck structure is used for network reconstruction, which adopts a cascading design: first, through 1×1 pointwise convolution (Pointwise Conv) and h-Swish activation, then down-sampling through stride 2 deep convolution (Deepwise Conv), then introducing SE attention mechanism to strengthen key features, and finally reducing dimension through pointwise convolution (Pointwise Conv); as shown in Figure 3 (a), the feature extraction module retains the residual connection, the left branch maintains the deep convolution with stride=1, and finally the feature fusion is performed with the shortcut branch. In particular, when down-sampling, the residual connection is cancelled to maintain the feature dimension matching, and all modules use h-Swish activation function to balance the calculation efficiency and nonlinear expression ability. Figure 3 (b) is a down-sampling module structure in a lightweight convolutional neural network, which forces down-sampling by stride=2 and removes the ADD residual connection, and through depth separable convolution, SE attention, and h-Swish, the model can efficiently process image features on mobile terminals.

[0098] The main network is reconstructed by introducing MobileNetV3 bottleneck, which significantly optimizes the model performance: on the one hand, the application of depth separable convolution and reverse residual structure effectively reduces parameter redundancy and computational complexity, and reduces model complexity; on the other hand, the introduction of SE attention mechanism enhances the network's ability to extract key features, improves the model's representation efficiency, and SPPF module is used for multi-scale feature extraction to enhance the model's detection ability for targets of different sizes. Figure 4 The improved lightweight backbone network structure is shown, which realizes a significant improvement in computational efficiency without losing 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, an improved network model YOLOv5s-CA is obtained, as shown in Figure 5 .

[0100] BiFPN structure is used instead of the original PANet, and the structure comparison is shown in Figure 6 . As shown in Figure 6 (a), Figure 6 (b), BiFPN is a weighted bidirectional feature pyramid network that removes high-level isolated nodes to simplify the network and introduces weighted connections between input and output nodes in the same layer, as shown by the arrows in the figure, which realizes the learning of feature importance distribution and significantly improves the feature fusion efficiency.

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

[0102] ; (1)

[0103] ; (2)

[0104] wherein, represents the intermediate fusion feature, Conv represents the convolution operation, , represents the learned trainable parameters, Resize represents the size adjustment operation, represents the input feature map, represents the deeper input feature map, represent the final output feature, , , represent the trainable parameters, represent the cross-layer feature, represent the minimum value, prevent the denominator from being zero, cause the calculation to collapse.

[0105] Step3: Improvement of the head network based on the loss function, the loss function adopts CIOU loss function, as shown in Figure 7 .

[0106] CIOU is defined as:

[0107] ;(3)

[0108] CIOU loss function is:

[0109] ;(4)

[0110] wherein, is the weight coefficient, balancing the influence of the aspect ratio term, is the difference between the predicted box and the real box aspect ratio.

[0111] EIOU loss function makes key improvements on the basis of CIOU, and improves the positioning accuracy by introducing a more refined aspect ratio optimization mechanism. As shown in Figure 8 .

[0112] EIOU is defined as:

[0113] ;(5)

[0114] EIOU loss function is:

[0115] ;(6)

[0116] wherein, b and are the center coordinates of the anchor frame and the real frame, denotes the Euclidean distance between the two, c is the diagonal distance of the minimum bounding rectangle, and are the minimum bounding rectangle width and height, , are the labeled frame width and height; and are the anchor frame width and height.

[0117] For the challenge of target detection in complex scenes, EIOU loss function is used instead of the original CIOU loss, and its excellent aspect ratio adaptability can effectively improve the detection accuracy of random size targets.

[0118] In the embodiment, multi-angle image cooperative image stitching:

[0119] The present application constructs a global panoramic perception view by multi-view image stitching and distributed target fusion technology, effectively eliminates the single robot visual field blind area, and expands the monitoring coverage. At the same time, by using the complementary characteristics of multi-robot view, the problem of tracking interruption caused by target occlusion is solved, and the robustness and tracking continuity of the system in dynamic environment are improved. The following steps are adopted to realize it:

[0120] Step 1: In the image preprocessing stage, the present application introduces an advanced color transfer algorithm, which realizes the color consistency correction between multi-view images by establishing the 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 linear transformation relationship is established as shown in formula (7).

[0121] (7)

[0122] In the image conversion process shown in formula (7), the source image and the target image are composed of three color channels. Specifically, the three-channel original data of the source image is represented as (R, G, B), and its corresponding mean and standard deviation are (R_mean, G_mean, B_mean) and (R_std, G_std, B_std); similarly, the output data of the target image is (L, A, B), and its statistical characteristic parameters include mean (L_mean, A_mean, B_mean) and standard deviation (L_std, A_std, B_std).

[0123] Step 2: One of the bottlenecks of the calculation efficiency of ORB algorithm is that the FAST feature point detection mechanism it uses needs to traverse the search of all image pixels. For the optimization problem of the calculation efficiency of FAST feature detection algorithm, an improved method based on image region screening is proposed. Since the corner points detected by FAST algorithm are usually distributed in the regions with significant changes in pixel values, the quadtree decomposition technology is introduced, and by analyzing the image gray distribution characteristics, the regions with gentle gray changes in the reference image and the image to be registered are automatically filtered out. This region screening mechanism effectively reduces the search range of feature point detection, and only in the regions with rich texture, FAST detection is performed, thereby significantly reducing the time consumption in the feature extraction stage.

[0124] ​​​​​​​​​​​​​​​Recursive quadtree partitioning is applied to the original image region S, and the center point coordinates of each sub-region are accurately located in each partitioning process, as shown in Figure 9

[0125] Determining the quadtree decomposition depth. In order to prevent the quadtree partitioning from being too deep and introducing additional time consumption, the depth of the quadtree partitioning needs to be determined.

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

[0127] For the retained target region , the region center point p1 is first determined, and then two symmetric points (p2-p5) are selected in the vertical and horizontal directions, and the distance between these sampling points and the center point is 1 / 2 of the region width or height (it has been verified through experiments that this ratio has the best effect). Based on these five sampling points, the region feature value is calculated by formula (8), wherein represents the pixel value of the jth sampling point in the ith sub-region, represents the feature value of the ith sub-region.

[0128] ; (8)

[0129] In the region screening stage, the Manhattan distance is used as the similarity measurement index. As shown in formula (9), wherein D i represents the Manhattan distance between the pixel points in the ith sub-region. By setting a threshold (z is valued according to the specific application scenario to distinguish the regions with uniform and non-uniform pixel distribution, and realize the region screening effect), the region screening is performed: when , it is determined that the pixel distribution of the region is uniform, and it is excluded; otherwise, it is retained and further decomposed.

[0130] ; (9)

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

[0132] Figure 10 A search domain screening schematic diagram is shown. This process adopts a quadtree decomposition strategy, wherein the T mark indicates that the pixel gray scale of the region changes significantly and needs to be further decomposed, and the F mark indicates that the region gray scale distribution is uniform and is excluded. The algorithm execution process is as follows: the original image S is used as the root node of the quadtree for initial partitioning, the gray scale change of each sub-region is evaluated, the regions that meet the conditions are recursively decomposed, and finally the output is as shown in Figure 10 ​the optimal search region indicated by d in the figure.

[0133] Step3: Under the condition of fixed neighborhood point number N, the maximum threshold value that keeps pixel point P still able to be determined as a corner point is defined as the corner response value of the point. For each candidate pixel point P, the following operations are performed: the absolute difference of gray scale of P and 8 adjacent pixel points on the circumference is calculated to obtain the minimum value d; considering the extension pixel points adjacent to the head and tail, the absolute difference of gray scale of each extension pixel point and P is calculated and ; through the comparison operation shown in formula (10), the corner response value of the current iteration is determined .

[0134] ; (10)

[0135] After complete traversal calculation of the 16 sampling points on the circumference, the corner response value of pixel point P is finally determined as:

[0136] ; (11)

[0137] After obtaining the corner response value of each pixel point, the response value comparison is performed on each candidate point in its 3*3 neighborhood range, and only those points with the maximum response value in their 8-neighborhood are retained as the final corner points.

[0138] Step4: In order 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 index. First, for each corner point detected by FAST, the gradient values of all pixel points in the x and y directions are calculated in its neighborhood range using the Sobel operator, denoted as and ; then, the gradient combination quantities a, b, c are calculated according to formulas (13)-(15); finally, the response intensity of each corner point is obtained through the Harris response function shown in formula (12), where k is an empirical coefficient and s is a scale parameter. This fusion strategy not only retains the efficiency of the FAST algorithm, but also 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 represented by an offset vector to indicate the intensity. The ORB algorithm uses the intensity centroid method to determine the corner direction. First, the spatial moments of the image patch are calculated to establish the direction reference. The moments reflect the local gray distribution characteristics, as shown in equation (16).

[0144] (16)

[0145] In the formula , I(p, q) represents the gray value of the pixel point at (p, q), and p, q represent the order of the moment. The coordinates of the centroid of the image patch can be determined by calculating the first-order moment

[0146] (17)

[0147] In the formula , M0 represents the zero-order spatial moment, , M1x represents the first-order spatial moment in the x direction, , and M1y represents the first-order spatial moment in the y direction. The local neighborhood is defined with the corner as the center, and the main direction of the feature point is directly determined by calculating the spatial distribution characteristics of the region

[0148] (18)

[0149] To overcome the limitation of the FAST algorithm in matching feature points when processing scaled images, the concept of scale pyramid is introduced. First, a scaling factor is set, which is usually 1.2, and the number of layers L of the pyramid is generally selected as 8. Based on these parameters, the original image is gradually scaled down to create a series of scaled image sets. The scale of the scaled image is shown in equation (19):

[0150] (19)

[0151] , where is the scale of the original image, , and the value range of .

[0152] A multi-scale strategy is used to extract feature points, and the oFAST features are detected at each scale layer through the pyramid model and then fused to ensure scale invariance.

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

[0154] ​​​Step5: After obtaining improved ORB feature points, FLANN is used for fast nearest neighbor matching, and PROSAC algorithm is used to remove false matching points.

[0155] The FLANN algorithm library processes large-scale data matching problems through an efficient approximate nearest neighbor search technology. Figure 11 As shown in the figure, the implementation steps are as follows:

[0156] The classic PROSAC needs to sort the data points according to the similarity evaluation first, and the accuracy and rationality of the sorting will directly affect the subsequent sampling and model estimation. The classic PROSAC needs to deal with the definition and growth of the hypothesis generation set, and cannot rely too much on the evaluation function pre-sorting, nor can it treat all data points equally. In actual operation, it is difficult to reasonably control the hypothesis generation set, which is easy to lead to model estimation deviation or affect the calculation efficiency. The classic PROSAC, when combined with new feature extraction and matching algorithms (such as deep feature matching), has insufficient adaptability and synergy, and it is difficult to fully play the combined advantages.

[0157] The flow chart of the improved PROSAC algorithm is shown in the figure Figure 12 , and the specific implementation steps are as follows:

[0158] Step1: The feature point pair set obtained by FLANN rough matching is , and the ordered feature point pair set is generated in descending order according to the matching quality function .

[0159] ; (20)

[0160] Among them, represents the matching quality score of the i-th feature point pair, indicates the feature point pair.

[0161] Step2: Select the first n high-quality matching points from to form a set N, and define the quality standard of set N as shown in formula (21):

[0162] ; (21)

[0163] Step3: The feature point pairs are divided into 4-point groups in descending order of quality, the quality sum of each group is calculated and sorted, and the 4 groups of matching points with the highest quality are selected to calculate the homography matrix H:

[0164] ; (22)

[0165] Among them​ and are the matched point coordinates.

[0166] Step4: Compute the projection error of the remaining points using the homography H.

[0167] The geometric error between a feature point and its projected point is computed to evaluate the matching quality. If the error is less than the projection error threshold , it is considered as an inlier, otherwise as an outlier. Suppose the projected point of point is , then the projection error is computed as shown in equation (23).

[0168] (23)

[0169] Step5: The projection error threshold is defined as:

[0170] (24)

[0171] where h is the height of the H image, w is the width of the image, and c is a constant, usually set to 0.1.

[0172] Step6: Compare the number of inliers with the preset threshold (i.e., the threshold defined in equation (24) above), if the condition is met, return the optimal model parameters, otherwise continue iteration until the maximum number of iterations is reached.

[0173] Step7: Homography transformation describes the perspective mapping relationship of the same plane under different angles. When the camera takes pictures of a plane scene from different angles, there is a geometric transformation between images that can be represented as a homography matrix H. Suppose the feature points in image and the feature in image are matched point pairs, and H is the homography transformation matrix, the relationship between the two feature points is as follows:

[0174] (25)

[0175] In the above equation, H is a 3x3 matrix. For each matched point pair , substitute it into the above equation to get the following equation:

[0176] (26)

[0177] Further transformation can be obtained as follows:

[0178] ; (27)

[0179] The above formula is written as The equation set is in the form of the following formula:

[0180] ; (28)

[0181] Homography matrix solving needs to determine 8 independent parameters. Since each pair of matching points can establish two equations, theoretically 4 groups of non-collinear matching points can solve a unique solution. After eliminating redundant parameters by constraint conditions, the optimal homography matrix is calculated by using the least square method.

[0182] Step 8: The fade-in and fade-out fusion algorithm realizes smooth transition by dynamically adjusting the fusion weight of the overlapping area pixels. The algorithm calculates the mixing ratio based on the distance relationship between the pixel and the image boundary. The closer to the boundary, the higher the weight of the pixel, so that the splicing area presents a natural fade-out effect. The specific expression is as follows:

[0183] ; (29)

[0184] In formula (29), represents the pixel value of the output image after fusion at coordinates (x, y), represents the pixel value of the left image at coordinates (x, y), represents the pixel value of the right image at coordinates (x, y), represents the fusion weight of the left image at coordinates (x, y), represents the fusion weight of the right image at coordinates (x, y), and respectively represent the non-overlapping areas on the left and right sides, represents the overlapping area. The pixel value of the output image is composed of three parts: the non-overlapping area directly uses the original pixel value, and the overlapping area calculates the mixing weight according to the pixel position. The weight coefficient is determined by the normalized distance of the pixel to the boundary. The weight of the left image decreases with the increase of the right boundary distance, and the weight of the right image decreases with the increase of the left boundary distance, and the sum of the two is always 1.

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

[0186] To solve the target occlusion problem in the following process, a time series convolution network TCN is introduced on the basis of the existing personnel identification network. Compared with the recurrent neural network structure such as LSTM and GRU, the time series convolution network shows significant advantages in time series modeling.

[0187] The construction process of the target recapture scheme based on TCN is as follows: Figure 13As shown, the mechanism integrates event-triggered judgment and overlap rate calculation functions. When the system detects a target loss event, the recapture process is automatically activated to realize target relocation by analyzing historical motion trajectory and real-time appearance features.

[0188] The application adopts TCN to construct a target re-identification mechanism, fully utilizes its time sequence memory characteristics, and realizes double verification by fusing target motion trajectory and appearance features. The following steps are adopted to realize it:

[0189] Step 1: prediction output, network input It is composed of the position change amount of the target bounding box at adjacent time, specifically the geometric difference amount of the detection box at the current time and the previous time. This time sequence difference feature effectively captures the change mode of the target motion trajectory.

[0190] ;(30)

[0191] The network input contains the horizontal and vertical position offset of the target in the image coordinate system, and the change amount of the detection box length and width. Through the layer-by-layer feature extraction of the time sequence convolution layer, the network finally outputs the future multi-step bounding box position prediction value and the scale change amount, forming a complete motion trajectory prediction result.

[0192] Among them, is the Y-axis coordinate at time t, is the X-axis coordinate at time t, is the width of the detection box at time t, is the height of the detection box at time t, and are the change amount of the target in the X-axis and Y-axis of the image coordinate system at time t, and are the change amount of the width and height of the detection box at time t. Through the layer-by-layer feature extraction of the TCN, the network finally outputs the future multi-step bounding box position prediction value and the scale change amount, forming a complete motion trajectory prediction result, denoted as :

[0193] ;(31)

[0194] Among them, is the step length predicted by the network, is the motion trajectory prediction result at time t;

[0195] Step 2: event-triggering mechanism, in the normal tracking state, the system ignores the prediction output of the time sequence network, keeps the basic classification process running, and maintains a fixed length ​Timing sliding window When the target loss event is detected, the re-identification verification mechanism is automatically activated, and the motion trajectory prediction result is cross-verified with the appearance classification output.

[0196] (32)

[0197] wherein, is the motion trajectory prediction result at the moment;

[0198] Step3: overlap rate calculation, re-identification verification link introduces confidence score index and The calculation formula is as follows:

[0199] (33)

[0200] (34)

[0201] By comparing the overlap degree of the bounding box output by the detection network and the re-identification prediction box , decision judgment is made in combination with the preset threshold value. When the comprehensive confidence evaluation result exceeds the set threshold value, it is determined that the target re-identification is successful, and the predicted bounding box compensates for the missing area of detection, otherwise it is considered as a matching failure. The threshold value is 1.7 obtained through multiple experiments.

[0202] (35)

[0203] In the embodiment, the multi-robot obstacle avoidance of potential field optimization is:

[0204] The present application introduces a random disturbance mechanism, which makes the robot get rid of the local potential energy trap by applying a random force. The present application introduces a random force , which provides additional power to help the robot escape when it is trapped. The random force expression is:

[0205] (36)

[0206] wherein: is the strength coefficient of the random force, which is used to control the size of the random disturbance; is a random vector, whose direction is uniformly distributed in , and the size is randomly generated in .

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

[0208] (37)

[0209] where, represents attractive force, represents repulsive force;

[0210] After introducing random disturbance, APF algorithm can apply random force to help robot escape from local minimum. As shown in Figure 14 and Figure 15 , the improved algorithm effectively solves the problem of local minimum.

[0211] On the other hand, when the influence range of obstacle contains target point, as the robot approaches the obstacle, the repulsive force increases and the attractive force decreases, resulting in the target being unreachable. The position relationship of the three is shown in Figure 16 .

[0212] When the robot, target point and obstacle are collinear and the target point is in the middle, the robot will be at a certain balance point, and the resultant force is 0. Therefore, the repulsive potential function is improved , as formula (38), so that the repulsive potential energy at the target point is 0.

[0213] ; (38)

[0214] where, represents order, ; represents that the target person is located in the repulsive potential field; is the repulsive gain coefficient; is the robot position; is the obstacle position; is the distance between the robot and the obstacle; is the farthest influence distance of the obstacle in the repulsive potential field; represents the target position; represents the distance between the robot and the target;

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

[0216] ; (39)

[0217] where, represents the total repulsive force vector received by the robot, represents the gradient of the repulsive potential field, represents the direct repulsive force of the obstacle, represents the corresponding inverse repulsive force associated with the target point.

[0218] where , is calculated as follows:

[0219] (40)

[0220] wherein: the direction of the robot is towards the target person.

[0221] (41)

[0222] wherein: the direction of the robot is towards the target person.

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

[0224] In this embodiment, the double closed-loop PID controller:

[0225] The present application adopts a parallel double closed-loop PID architecture, and precisely adjusts the position and direction angle through two independent control channels. The double closed-loop PID principle diagram is shown in Figure 19 .

[0226] In the position control channel, the outer ring PID generates a target linear velocity according to the deviation between the expected following distance and the actual distance , and the inner ring PID is responsible for accurate tracking of ; in the direction control channel, the outer ring PID calculates the target angular velocity by the deviation between the expected azimuth angle and the actual azimuth angle , and the inner ring PID realizes the rapid response of . The outputs of the two control channels are synthesized through the differential kinematics model: , wherein, is the linear velocity of the robot, obtained by the position ring PID output, is the angular velocity of the robot, obtained by the direction ring PID output, represents the left wheel linear velocity of the robot, represents the right wheel linear velocity of the robot, and W is the wheel distance, and finally generates four-wheel drive instructions.

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

[0228] (1) To verify the robustness of the target tracking algorithm proposed in the present application, a comprehensive evaluation was carried out on the self-made data stream, which covers a variety of complex scenes, including light changes, scale changes, occlusions, target leaving the field of view and background interference and other challenging conditions.

[0229] The experimental results show that the proposed method exhibits good tracking performance in different scenarios: in bright environments, it can effectively deal with complex backgrounds, similar pedestrian interference and other problems; in low light conditions, it can handle the challenge of blurred target features; even if the target temporarily leaves the single camera field of view or encounters pedestrian occlusion, it can still maintain stable tracking ability. Especially in scenarios with drastic changes in lighting conditions and severe occlusions, the method can still accurately detect and continuously track the target, showing strong environmental adaptability.

[0230] The experimental results show that the pedestrian following method proposed by the present application performs well in complex environments.

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

[0232] The present application builds an experimental environment based on the ROS Melodic and Gazebo 9.0.0 simulation platform, and selects Turtlebot3 burger as the core experimental platform. The robot adopts a compact design, with overall dimensions of 138mm x 178mm x 192mm (L x W x H), and has good mobility flexibility. Although Turtlebot3 burger adopts a two-wheel differential drive structure, its kinematic model is equivalent to that of a four-wheel differential drive robot, and both achieve steering control through the speed difference between the left and right wheels, so it can be used as an effective substitute platform for four-wheel differential robots to verify the algorithm. In terms of sensor configuration, the burger robot is equipped with a Kinect RGB-D camera and an LDS-01 laser radar sensor, of which the Kinect camera is responsible for target identification and feature extraction, and the LDS-01 laser radar is used for environment perception and obstacle detection, and the two work together to achieve accurate positioning and tracking of dynamic targets. The experiment deploys three Turtlebot3 burger robots in Gazebo, and uses the topic communication mechanism of ROS to realize information exchange between robots.

[0233] A 20m x 20m square simulation scene is constructed, and the site is space calibrated with a 1m x 1m grid. The obstacle settings in the environment include two types: static obstacles composed of cross-shaped inner walls, and other pedestrians as dynamic obstacles. To simulate the real tracking scene, the target personnel move according to the preset trajectory as shown in Figure 20 、 Figure 21 .

[0234] The system identification results are monitored in real time through the Rviz visualization tool, and the visualization effect is as shown in Figure 22 .

[0235] The driving trajectories of the three robots are as shown in Figure 23As shown, the results show that each robot well avoids the obstacles and achieves stable tracking of the target under the guidance of the tracking control strategy of the application.

[0236] The above-described embodiments are merely intended to describe the preferred modes of the application, and are not intended to limit the scope of the application. Various modifications and improvements to the technical solutions of the application made by those skilled in the art without departing from the design spirit of the application shall fall within the protection scope of the application as defined by the claims.

Claims

1. A multi-robot collaborative target recognition and tracking method in a dynamic environment, characterized by: The method comprises: Identify target personnel in real time; If the confidence level of the target person is lower than the threshold and there is a valid detection frame, 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, the multi-view collaborative mechanism is activated to build a global panoramic perception view through multi-view image stitching and distributed target fusion technology. At the same time, the complementary characteristics of multiple robot perspectives are used 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 obtained in real time, and the optimized APF and PID control strategies are used to establish a closed-loop system from target recognition to motion control to maintain the tracking strategy.

2. The method according to claim 1, characterized in that Methods for real-time identification of target individuals include: Based on the YOLOv5s model, the task of real-time target person identification is realized; Among them, 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 confidence level of the target person is lower than the threshold and a valid detection frame exists, the re-identification module is activated. The method of using the TCN network combined with time series features and motion patterns to confirm the identity includes: Prediction output: Network input It is composed of the position change of the target bounding box at adjacent moments, specifically expressed as the current moment With the previous moment The geometric difference of the detection box: ； The network input includes the horizontal and vertical position offsets of the target in the image coordinate system, as well as the changes in the length and width of the detection box. Through the layer-by-layer feature extraction of the temporal convolution layer, the network finally outputs the predicted position value and scale change of the bounding box for multiple steps in the future, forming a complete motion trajectory prediction result. in, is the Y-axis coordinate at time t, is the X-axis coordinate at time t, is the detection frame width at time t, is the height of the detection frame at time t, and They are The change of the target on the X-axis and Y-axis in the image coordinate system within a certain time. and They are The change in the width and height of the detection box within the moment is extracted layer by layer through TCN, and the predicted position value and scale change of the bounding box with the preset step size in the future are finally output to form a complete motion trajectory prediction result, which is recorded as : ； in, is the step size of network prediction, for Motion trajectory prediction results at each moment; Event trigger mechanism: In the normal tracking state, the system ignores the prediction output of the time series network, keeps the basic classification process running, and maintains a fixed length Time series sliding window ,When a target loss event is detected, the re-identification verification mechanism is activated to cross-validate the motion trajectory prediction results with the appearance classification output; ； in, for Motion trajectory prediction results at each moment; Overlap rate calculation: Introducing a credibility score indicator into the re-identification verification phase and , the calculation formula is as follows: ； ； By comparing the bounding boxes output by the detection network and re-identification prediction box The overlap degree is combined with the preset threshold to make a decision. When the comprehensive credibility evaluation result exceeds the set threshold, the target re-identification is judged to be successful, and the predicted bounding box is used to compensate for the missing detection area. Otherwise, it is considered a match failure. 。 4. The method according to claim 1, wherein If the target is lost, the multi-view collaborative 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 used to solve the tracking interruption problem caused by target occlusion. The following methods are used: In the image preprocessing stage, a color migration algorithm is introduced to establish a color mapping relationship between the source image and the target image; Recursively divide the original image area S into four equal parts, and locate the coordinates of the center point of each sub-block in each division process; Under the condition of a fixed number of neighborhood points N, the maximum threshold that can keep the pixel point P still being judged as a corner point is defined as the corner point response value of the point; Harris response value is introduced as an auxiliary evaluation index, the FAST algorithm is used to characterize the intensity of corner points through the offset vector, the ORB algorithm is used to determine the direction of corner points through the intensity centroid, and the concept of scale pyramid is introduced to achieve this 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 combined to eliminate mismatched points.

5. The method according to claim 4, characterized in that After obtaining the improved ORB feature points, FLANN is used for nearest neighbor matching. The methods of removing mismatched points by combining the PROSAC algorithm include: The feature point pair set obtained by FLANN rough matching , according to the matching quality function Sort in descending order to generate an ordered set of feature point pairs ; from In the example, the first n high-quality matching points are selected in sequence to form a set N, and the quality standard of the set N is defined; Divide the feature point pairs into groups of 4 points in descending order of quality, calculate the sum of the quality of each group and sort them, and select the 4 groups of matching points with the highest quality to calculate the homography matrix H; Use the homography matrix H to calculate the projection error of the remaining points; The geometric error between the feature point and the corresponding projection point is calculated to evaluate the matching quality. If the error is less than the threshold, the feature point is considered an internal point, otherwise it is considered an external point. Compare the number of inliers with the preset threshold. If the conditions are met, the optimal model parameters are returned. Otherwise, continue to iterate until the maximum number of iterations is reached.

6. The method according to claim 1, characterized in that If the confidence level of the target person is higher than the threshold, the target position is acquired in real time, and the optimized APF and PID control strategies are 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 position, the robot calculates the motion trajectory based on the distributed APF. The gravitational component guides the robot to converge to the target, while the repulsive component avoids obstacles. Using a hierarchical PID architecture, the outer-loop attitude controller converts the desired coordinates into speed commands, and the inner-loop speed controller tracks the target through encoder feedback to generate a differential drive signal.

7. The method according to claim 6, characterized in that After confirming the target position, the robot calculates the motion trajectory based on the distributed APF. The gravitational component guides the robot to converge toward the target, and the repulsive component avoids obstacles in the following ways: Introducing random forces , provides additional power when the robot is trapped. The expression of this random force is: ； in: is the intensity coefficient of the random force, which is used to control the magnitude of the random disturbance; is a random vector with a direction in Evenly distributed within the range, the size is Randomly generated within the range; After the introduction of random forces, the total force acting on the robot is expressed as the vector sum of gravity, repulsion, and random forces: ； 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 the preset equilibrium point and the net force will be 0. The repulsive potential field function is improved so that the repulsive potential energy at the target point is 0. The repulsive potential field function is: ； in, represents the order, ; Indicates that the target person is located in the repulsive potential field; is the repulsion gain coefficient; is the robot position; is the obstacle location; is the distance between the robot and the obstacle; is the farthest influence distance of the repulsive potential field of the obstacle; Indicates the target location; Indicates the distance between the robot and the target; The repulsive force exerted on the robot by the improved repulsive potential field function is: ； in, Represents the total repulsive force vector on the robot, represents the gradient of the repulsive potential field, represents the direct repulsion of the obstacle, Represents the reverse repulsive force associated with the corresponding target point.

8. The method according to claim 6, characterized in that Using a hierarchical PID architecture, the outer-loop attitude controller converts the desired coordinates into speed commands, and the inner-loop speed controller tracks the target through encoder feedback. The method for generating the differential drive signal includes: In the position control channel, the outer loop PID follows the desired distance Actual distance Deviation Generate target linear velocity The inner loop PID is responsible for Conduct tracking; In the direction control channel, the outer loop PID passes the desired azimuth angle and the actual azimuth Deviation Calculate target angular velocity , and is implemented by the inner loop PID Rapid response; The outputs of the two control channels are synthesized using a differential kinematics model: , ,in, is the robot linear velocity, obtained from the position loop PID output, is the robot angular velocity, which is obtained from the direction loop PID output. represents the robot's left wheel linear speed, represents the linear velocity of the robot's right wheel, W is the distance between the wheels, and finally generates a four-wheel drive command.

Images