Robot navigation decision-making method based on target recognition

Abstract

The present invention relates to the field of robot navigation technology, and in particular to a robot navigation decision method based on target recognition, comprising the following steps: S1, obtaining environmental images and image preprocessing: obtaining real-time image data of the robot's surrounding environment; S2, target recognition and feature extraction: identifying target objects in the environment; S3, environmental modeling: generating an environmental map; S4, path planning: calculating the optimal path of the robot from the current position to the target point; S5, path prediction and correction: dynamically adjusting and optimizing the robot's navigation path; S6, navigation decision and dynamic adjustment: guiding the robot to perform real-time navigation; S7, feedback mechanism and optimization: continuously optimizing target recognition and navigation decision. The present invention ensures that the robot can flexibly respond to obstacles and other emergencies in a complex dynamic environment, and significantly improves the reliability and safety of the navigation system.

Description

Technical Field

[0001] The present invention relates to the field of robot navigation technology, and in particular to a robot navigation decision method based on target recognition. Background Art

[0002] With the rapid development of artificial intelligence and robotics technology, robots are being used more and more widely in industry, medical care, services and other fields. Among them, robot autonomous navigation technology, as a key technology that enables robots to move and perform tasks autonomously, has received widespread attention. Autonomous navigation technology includes environmental perception, path planning, navigation control and other aspects, aiming to enable robots to move safely and efficiently in complex and changing environments.

[0003] Existing robot navigation technology mainly relies on pre-built environmental maps and fixed path planning algorithms, and usually uses sensors such as lidar and cameras for environmental perception. However, existing technologies have many shortcomings. First, the accuracy and real-time performance of environmental perception and target recognition are insufficient, and they are easily affected by environmental changes and noise. Second, the flexibility and adaptability of path planning algorithms are poor, and they have difficulty coping with a variety of obstacles in complex dynamic environments. Finally, the dynamic adjustment capability of navigation decisions is limited, and there is a lack of real-time optimization and feedback mechanisms, which cannot ensure that robots can always navigate efficiently and safely in a changing environment.

[0004] The purpose of the present invention is to provide a robot navigation decision-making method based on target recognition, which significantly improves the reliability and adaptability of the system and ensures that the robot can complete tasks safely and efficiently in complex environments. Summary of the invention

[0005] Based on the above objectives, the present invention provides a robot navigation decision method based on target recognition.

[0006] The robot navigation decision method based on target recognition includes the following steps:

[0007] S1: Use cameras and image sensors to obtain real-time image data of the robot's surrounding environment, and pre-process the acquired image data, including image denoising, grayscale conversion, and edge detection;

[0008] S2: Perform target recognition on the preprocessed image data, identify the target object in the environment, and extract feature information from the identified target object;

[0009] S3: Integrate feature information and use a grid map to construct the integrated feature information into a grid environment model to generate an environment map, which includes the location, shape and related information of the target object;

[0010] S4: Based on the constructed environment model, calculate the optimal path from the current position of the robot to the target point;

[0011] S5: Based on the results of path planning, a generative adversarial network is introduced to predict and correct the path, dynamically adjust and optimize the robot's navigation path, including:

[0012] S51: Generate potential paths using a generative adversarial network, evaluate the generated potential paths, and output the feasibility score of each generated path; specifically, the generator not only inputs random noise, but also inputs the current environment model, robot position information and path constraints to generate potential paths; the discriminator inputs the generated path and the real path, and evaluates them in combination with environmental features and multi-scale features, and outputs the feasibility score of the path; through adversarial training of the generator and the discriminator; the loss function of the generator is:

[0013] ;

[0014] in, is random noise, is the environmental feature vector, is the path constraint, is a multi-scale feature, is the probability distribution of noise, is the probability distribution of environmental characteristics, is the probability distribution of path constraints, D represents the identifier, It is expectation;

[0015] S52: According to the feasibility score result, select the path with the highest feasibility as the current navigation path;

[0016] S53: During the actual navigation process of the robot, the environment changes are monitored in real time, and the path is dynamically corrected through the generative adversarial network;

[0017] S6: Generate navigation instructions based on the results of path prediction and correction, combined with the robot's current position, speed and direction, to guide the robot to perform real-time navigation and dynamically adjust navigation decisions;

[0018] S7: Continuously optimize target recognition and navigation decisions through real-time image data and navigation result feedback.

[0019] Furthermore, the S1 step includes:

[0020] S11, image data acquisition: using a camera and an image sensor installed on the robot to acquire real-time image data of the robot's surroundings;

[0021] S12, image denoising: using a Gaussian filter to denoise the acquired image data;

[0022] S13, image grayscale processing: grayscale the denoised image data using a weighted average method to convert the color image into a grayscale image;

[0023] S14, edge detection processing: using the Canny edge detection algorithm to perform edge detection on the grayscale image to extract edge information in the image.

[0024] Furthermore, the S2 step includes:

[0025] S21, target recognition: Use the convolutional neural network (CNN) model to perform target recognition on the preprocessed image data to identify the target objects in the environment, including obstacles, target points and navigation signs;

[0026] S22, feature extraction: extracting feature information from the identified target object, including the shape, size, position and motion state of the target object, specifically including:

[0027] S221, shape feature extraction: using the Canny edge detection algorithm to extract the contour and edge of the target object, and calculate the shape features, including perimeter, area and contour approximation polygon;

[0028] S222, size feature extraction: calculate the bounding box of the target object to obtain its width, height and area;

[0029] S223, position feature extraction: determine the position of the target object in the image and calculate its centroid coordinates;

[0030] S224, motion state feature extraction: calculate the motion trajectory and speed of the target object through continuous multi-frame image data, and estimate the motion vector of the target object using an optical flow algorithm.

[0031] Furthermore, the S3 step specifically includes:

[0032] S31, feature information integration: integrating the feature information of the target object obtained from target recognition and feature extraction;

[0033] S32, environmental model construction: Based on the integrated feature information, a grid environment model is constructed using a grid map method; the grid map divides the environment into a number of grids, each grid contains a probability value, indicating the possibility of the corresponding area being occupied, and the occupancy probability of the grid is updated using the position and size information of the target object to generate a grid map containing the target object; the mathematical formula of the grid environment model is:

[0034] ;

[0035] in, It is The probability that a grid is occupied is is the log-odds of the grid, From time 1 to The sensor measurement value;

[0036] S33, environment map generation: converting the constructed environment model into an environment map.

[0037] Further, the S4 step includes:

[0038] S41, path planning: Based on the constructed grid environment model, the A* algorithm is used to calculate the optimal path from the current position of the robot to the target point;

[0039] S42, path optimization: In the path planning process, the robot size and motion capability are considered, and the path is smoothed using the spline interpolation method. The calculation formula is:

[0040] ;

[0041] in, is a point on the smooth path, is the control point, as a parameter.

[0042] Further, the step S51 includes:

[0043] S511, Generator: The generator not only inputs random noise, but also inputs the current environment model, robot position information and path constraints to generate potential paths. The calculation formula is:

[0044] ;

[0045] in, is random noise, is the environmental feature vector (including the current environmental model and robot position information), are path constraints (such as maximum turning angle, minimum safe distance, etc.), is the weight matrix of the generator, is the first bias, is the activation function;

[0046] S512, discriminator: The discriminator inputs the generated path and the real path, and evaluates them in combination with environmental features and multi-scale features, and outputs the feasibility score of the path. The calculation formula is:

[0047] ;

[0048] in, is the input path (including generated path and real path), is the environmental feature vector, is a multi-scale feature (such as local path features and global path features), is the weight matrix of the discriminator, is the second bias, is the activation function;

[0049] S513, Generate adversarial training: Through adversarial training between the generator and the discriminator, the calculation formula is:

[0050] The loss function of the generator is:

[0051] ;

[0052] in, is random noise, is the environmental feature vector, is the path constraint, is a multi-scale feature, is the probability distribution of noise, is the probability distribution of environmental characteristics, is the probability distribution of path constraints, It is expectation;

[0053] The loss function of the discriminator is:

[0054] ;

[0055] in, is the real path data, is the environmental feature vector, is a multi-scale feature, is the probability distribution of the real path data, is the probability distribution of environmental characteristics, is the probability distribution of multi-scale features;

[0056] S514, Path Generation and Evaluation:

[0057] Path generation: The generator inputs random noise, environmental characteristics and path constraints to generate potential paths. The calculation formula is:

[0058] ;

[0059] Path evaluation: The discriminator inputs the generated path, environmental features, and multi-scale features, and outputs the feasibility score of the path. The calculation formula is:

[0060] ;

[0061] True path evaluation: The discriminator inputs the true path, environmental features, and multi-scale features, and outputs the feasibility score of the true path. The calculation formula is:

[0062] .

[0063] Further, the step S52 includes:

[0064] S521, path scoring sorting: sort all generated potential paths Feasibility score Sort to get a list of paths sorted in descending order by score , corresponding to , the calculation formula is:

[0065] ;

[0066] in, For the discriminator path The feasibility score, is the environmental feature vector, It is a multi-scale feature;

[0067] S522, select the best path: select the path with the highest feasibility score from the sorted path list As the current navigation path, the calculation formula is:

[0068] ;

[0069] in, The path with the highest feasibility score is Give it the corresponding score.

[0070] Furthermore, the step S53 includes:

[0071] S531, Environmental Change Monitoring: Use cameras and sensors to obtain the latest image data and sensor data of the robot's surroundings in real time, and update the environmental feature vector , the calculation formula is:

[0072] ;

[0073] in, and are the latest image data and sensor data, respectively. Extract function for environmental features;

[0074] S532, Path Regeneration: Generator Input Random Noise , the updated environmental feature vector and path constraints , generating new potential paths , the calculation formula is:

[0075] ;

[0076] in, is the generator function, is the weight matrix of the generator, is the first bias, is the activation function;

[0077] S533, Path re-evaluation: The discriminator inputs a new potential path , the updated environmental feature vector and multi-scale features , output the new path feasibility score , the calculation formula is:

[0078] ;

[0079] in, is the discriminator function, is the weight matrix of the discriminator, is the second bias, is the activation function;

[0080] S534, Path selection and switching: score the new path based on its feasibility , choose the path with the highest feasibility score As the current navigation path, the robot navigation path is switched in real time. The calculation formula is:

[0081] .

[0082] Further, the S6 step includes:

[0083] S61, Navigation decision: Generate navigation instructions based on the results of path prediction and correction, combined with the robot's current position, speed and direction, to guide the robot to perform real-time navigation, including:

[0084] S611, path information acquisition: obtain the currently selected best path from the path correction ;

[0085] S612, status information update: obtain the current position of the robot in real time ,speed and direction , the calculation formula is:

[0086] ;

[0087] S613, generating navigation instructions: according to the best path and the current state of the robot, and generates navigation instructions using the model predictive control (MPC) algorithm , guiding the robot to perform real-time navigation, the calculation formula is:

[0088] ;

[0089] in, For control inputs (such as speed and steering angle), and is the weight matrix, is the Euclidean distance;

[0090] S62, dynamic adjustment: During the robot navigation process, the environment changes and robot state changes are monitored in real time, and navigation decisions are adjusted dynamically.

[0091] . Further, the S7 step includes:

[0092] S71, Feedback Mechanism: Establish a feedback mechanism through real-time image data and navigation results to continuously optimize target recognition and navigation decisions, including:

[0093] S711, Real-time Data Acquisition: Use cameras and sensors to acquire image data and sensor data in real time during robot navigation;

[0094] S712, Navigation result recording: Record the path selection, obstacle avoidance and navigation effect of the robot during navigation, including the actual path , obstacle avoidance points and time to reach the target ;

[0095] S713, data feedback: Feedback the real-time acquired image data, sensor data and navigation results. The calculation formula is:

[0096] ;

[0097] in, is real-time image data, is real-time sensor data, is the navigation result;

[0098] S72, target recognition optimization: Based on the results of the feedback mechanism, the weights of the convolutional neural network (CNN) model are updated using an online learning algorithm. The calculation formula is:

[0099] ;

[0100] in, are model parameters, is the learning rate, is the loss function, is the gradient calculation;

[0101] S73, Navigation decision optimization: According to the results of the feedback mechanism, the navigation strategy is updated using the reinforcement learning algorithm. The calculation formula is:

[0102] ;

[0103] in, is the current state, is the action, s' is the state at the next moment, a' is the action at the next moment, For instant rewards, is the discount factor, is the learning rate.

[0104] Beneficial effects of the present invention:

[0105] The present invention significantly improves the navigation accuracy, real-time and adaptability of robots in complex environments through a robot navigation decision-making method based on target recognition. It uses cameras and sensors to obtain environmental image data in real time, and through efficient image preprocessing, target recognition and feature extraction methods, it accurately identifies target objects in the environment and extracts their detailed feature information. These feature information are used to build accurate environmental models and environmental maps, providing a solid foundation for subsequent path planning and navigation decisions.

[0106] The present invention introduces a generative adversarial network to predict and correct the path. The generator generates a potential path in combination with environmental features and path constraints, and the discriminator scores the feasibility of the generated path to ensure the selection of the optimal path. This not only improves the flexibility and adaptability of path planning, but also ensures that the robot can flexibly respond to obstacles and other emergencies in a complex dynamic environment by real-time monitoring of environmental changes and dynamic correction of the path, significantly improving the reliability and safety of the navigation system. In the multi-source data input step, the generator introduces environmental features (including the current environmental model, robot position information) and complex path constraint information to enhance the ability of the GAN model to generate potential paths. Multi-scale features are introduced: by introducing local and global multi-scale features, the discriminator is more comprehensive in evaluating the feasibility of the path, and can simultaneously consider local path details and global path planning. The addition of path constraints: path constraints (such as maximum turning angle, minimum safe distance, etc.) are added to ensure that the generated path is not only feasible, but also meets the safety requirements in practical applications. Compared with the simple rules of traditional path planning methods, path constraints are used as an intrinsic part of the model, which greatly improves the rationality and safety of path planning.

[0107] The present invention generates navigation instructions by combining a model predictive control algorithm, monitors the changes in the state of the environment and the robot in real time during the navigation process, and dynamically adjusts the navigation decision to ensure that the robot can smoothly reach the target point. In addition, the target recognition and navigation decision models are continuously optimized by using real-time image data and feedback from navigation results, and the model parameters are dynamically updated using online learning and reinforcement learning algorithms, thereby improving the overall performance and robustness of the system and ensuring that the robot always maintains efficient and safe navigation capabilities in a constantly changing environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0109] Figure 1 A schematic diagram of a decision-making method flow chart of an embodiment of the present invention;

[0110] Figure 2 Schematic diagram of path prediction and correction according to an embodiment of the present invention.

[0111] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0112] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0113] like Figure 1-Figure 2 As shown, the robot navigation decision method based on target recognition includes the following steps:

[0114] S1, obtaining environmental images and image preprocessing: using cameras and image sensors to obtain real-time image data of the robot's surrounding environment, and preprocessing the acquired image data, including image denoising, grayscale conversion, and edge detection, to improve image quality and the efficiency of subsequent processing;

[0115] S2, target recognition and feature extraction: perform target recognition on the preprocessed image data, identify the target objects in the environment, and extract feature information from the identified target objects;

[0116] S3, environment modeling: Use the extracted feature information to build an environment model and generate an environment map, which includes the location, shape and related information of the target object;

[0117] S4, path planning: based on the constructed environment model, calculate the optimal path from the current position of the robot to the target point;

[0118] S5, path prediction and correction: Based on the results of path planning, a generative adversarial network (GAN) is introduced to perform path prediction and correction, dynamically adjust and optimize the robot's navigation path, including:

[0119] S51, Path Generation and Evaluation: Generate potential paths using a generative adversarial network (GAN), evaluate the generated potential paths, and output the feasibility score of each generated path;

[0120] S52, path selection: according to the result of feasibility scoring, select the path with the highest feasibility as the current navigation path;

[0121] S53, path correction: During the actual navigation process of the robot, the environment changes are monitored in real time, and the path is dynamically corrected through the generative adversarial network;

[0122] S6, Navigation decision and dynamic adjustment: Based on the results of path prediction and correction, combined with the robot's current position, speed and direction, generate navigation instructions to guide the robot to navigate in real time and dynamically adjust navigation decisions to ensure that the robot can adapt to environmental changes, avoid obstacles and reach the target point smoothly;

[0123] S7, Feedback mechanism and optimization: Through the feedback of real-time image data and navigation results, the target recognition and navigation decision are continuously optimized to improve the overall performance and reliability of the navigation system;

[0124] Through the above steps, the robot's navigation accuracy, real-time and adaptability in complex environments are significantly improved, ensuring the safety of the navigation process and the reliability of the overall system.

[0125] Acquisition of environmental images and image preprocessing in S1 include:

[0126] S11, image data acquisition: using a camera and an image sensor installed on the robot to acquire real-time image data of the robot's surroundings;

[0127] S12, image denoising: use a Gaussian filter to denoise the acquired image data to reduce noise interference in the image and improve the clarity of the image. The calculation formula is:

[0128] ;

[0129] in, is the original image, is the denoised image, is the standard deviation of the Gaussian filter;

[0130] S13, image grayscale processing: Use the weighted average method to grayscale the denoised image data, and convert the color image into a grayscale image. The calculation formula is:

[0131] ;

[0132] in, , and are the pixel values ​​of the red, green, and blue channels of the color image, respectively. is the pixel value of the grayscale image, , and are the weight coefficients for the red, green, and blue channels respectively;

[0133] S14, edge detection processing: Use the Canny edge detection algorithm to perform edge detection on the grayscale image and extract edge information in the image for subsequent target recognition and feature extraction, specifically including:

[0134] S141, calculate the gradient of the image: use the Sobel operator to calculate the gradient of the image in the horizontal and vertical directions. The calculation formula is:

[0135] ;

[0136] ;

[0137] S142, calculate the gradient amplitude and direction:

[0138] ;

[0139] ;

[0140] S143, non-maximum suppression: suppress non-maximum pixels along the gradient direction to refine the edge;

[0141] S144, dual threshold detection and edge connection: use high and low dual thresholds to perform edge detection and connection to generate a final binary edge map;

[0142] Through the above steps, the quality and efficiency of image preprocessing are improved, ensuring the accuracy of subsequent target recognition and navigation decisions.

[0143] Target recognition and feature extraction in S2 include:

[0144] S21, target recognition: Use the convolutional neural network (CNN) model to perform target recognition on the preprocessed image data to identify the target objects in the environment, including obstacles, target points and navigation signs;

[0145] The Convolutional Neural Network (CNN) model consists of:

[0146] Convolutional layer: The calculation formula of convolution operation is:

[0147] ;

[0148] in, is the output feature map at position No. The value of the channel, is the input image at position No. The value of the channel, is the weight of the convolution kernel, is the bias, is the ReLU activation function;

[0149] Pooling layer: The pooling operation uses maximum pooling, and the calculation formula is:

[0150] ;

[0151] in, The output feature map of the pooling layer is at position No. The value of the channel, is the input value within the pooling window, is the pooling step size;

[0152] Fully connected layer: The calculation formula of the fully connected layer is:

[0153] ;

[0154] in, is the output vector values, is the first values, is the weight, is the third bias, is the ReLU activation function;

[0155] S22, feature extraction: extracting feature information from the identified target object, including the shape, size, position and motion state of the target object, specifically including:

[0156] S221, shape feature extraction: Use the Canny edge detection algorithm to extract the contour and edge of the target object, and calculate the shape features, including the perimeter, area, and contour approximate polygon. The calculation formula is:

[0157] ;

[0158] in, is the circumference of the target object, is the edge point coordinate;

[0159] S222, size feature extraction: Calculate the bounding box of the target object to obtain its width, height and area. The calculation formula is:

[0160] ;

[0161] in, is the area, is the width, is the height;

[0162] S223, position feature extraction: determine the position of the target object in the image and calculate its centroid coordinates. The calculation formula is:

[0163] ;

[0164] in, is the centroid coordinate, is the pixel coordinate of the target object;

[0165] S224, motion state feature extraction: calculate the motion trajectory and speed of the target object through continuous multi-frame image data, and estimate the motion vector of the target object using the optical flow algorithm. The calculation formula is:

[0166] ;

[0167] in, For speed, and The optical flow is and Directional component;

[0168] Through the above steps, high-precision recognition and detailed feature analysis of target objects in the environment are achieved, which not only improves the accuracy and robustness of target recognition, but also provides rich basic data for subsequent environmental modeling and path planning by extracting multi-dimensional features such as the shape, size, position and motion state of the target objects, significantly improving the robot's autonomous navigation capability and decision-making efficiency in complex environments.

[0169] Environmental modeling in S3 includes:

[0170] S31, feature information integration: integrating the feature information of the target object obtained from target recognition and feature extraction;

[0171] S32, environmental model construction: based on the integrated feature information, a grid environmental model is constructed using a grid map method;

[0172] The grid map divides the environment into several grids. Each grid contains a probability value, which indicates the possibility of the corresponding area being occupied. The occupancy probability of the grid is updated using the position and size information of the target object to generate a grid map containing the target object. The calculation formula is:

[0173] ;

[0174] in, It is The probability that a grid is occupied is is the log-odds of the grid, From time 1 to The sensor measurement value;

[0175] S33, environment map generation: converting the constructed environment model into an environment map, the environment map includes the location, shape and related information of the target object, which is used for subsequent path planning and navigation decisions;

[0176] Through the above content, a detailed environmental model is constructed and an environmental map containing rich information is generated, which improves the robot's perception of the environment and provides a solid foundation for accurate path planning and efficient navigation.

[0177] Path planning in S4 includes:

[0178] S41, path planning: Based on the constructed grid environment model, the A* algorithm is used to calculate the optimal path from the current position of the robot to the target point, including:

[0179] S411, Heuristic function: Define the Euclidean distance estimate from the node The cost to the target node is calculated as:

[0180] ;

[0181] in, For Node The coordinates of is the coordinate of the target node;

[0182] S412, Cost Function: Define the total cost function , from the starting point to the node The actual cost and the heuristic cost The calculation formula is:

[0183] ;

[0184] S413, node expansion: starting from the starting point, add the starting point node to the open list and calculate the number of its adjacent nodes. value, and add the adjacent nodes to the open list, select The node with the smallest value is expanded and its adjacent nodes are updated. and value, moves the expanded node to the closed list;

[0185] S414, path search: repeat node expansion until the target node is expanded or the open list is empty. If the target node is expanded, trace back the connections between the nodes to generate the optimal path from the starting point to the target point;

[0186] S42, path optimization: In the process of path planning, the robot size and motion ability are taken into consideration, and the spline interpolation method is used to smooth the path to avoid sharp turns and narrow passages, thereby improving the feasibility and safety of the path. The calculation formula is:

[0187] ;

[0188] in, is a point on the smooth path, is the control point, is a parameter with a range of ;

[0189] Through the above steps, the optimal path from the current position of the robot to the target point can be effectively calculated. At the same time, the path is smoothed by the spline interpolation method, which improves the feasibility and safety of the path, ensures that the robot can successfully avoid obstacles and navigate with the optimal path in complex environments, and significantly improves the accuracy and efficiency of the navigation system.

[0190] The Generative Adversarial Network (GAN) in S51 includes:

[0191] S511, Generator: The generator not only inputs random noise, but also inputs the current environment model, robot position information and path constraints to generate potential paths. The calculation formula is:

[0192] ;

[0193] in, is random noise, is the environmental feature vector (including the current environmental model and robot position information), are path constraints (such as maximum turning angle, minimum safe distance, etc.), is the weight matrix of the generator, is the first bias, is the activation function;

[0194] S512, discriminator: The discriminator inputs the generated path and the real path, and evaluates them in combination with environmental features and multi-scale features, and outputs the feasibility score of the path. The calculation formula is:

[0195] ;

[0196] in, is the input path (including generated path and real path), is the environmental feature vector, is a multi-scale feature (such as local path features and global path features), is the weight matrix of the discriminator, is the second bias, is the activation function;

[0197] S513, Generative Adversarial Training: Through adversarial training between the generator and the discriminator, the generator can generate potential paths that are more adaptable to the environment and path constraints, while the discriminator can more accurately evaluate the feasibility of the path. The calculation formula is:

[0198] The loss function of the generator is:

[0199] ;

[0200] in, is random noise, is the environmental feature vector, is the path constraint, is a multi-scale feature, is the probability distribution of noise, is the probability distribution of environmental characteristics, is the probability distribution of path constraints, It is expectation;

[0201] The loss function of the discriminator is:

[0202] ;

[0203] in, is the real path data, is the environmental feature vector, is a multi-scale feature, is the probability distribution of the real path data, is the probability distribution of environmental characteristics, is the probability distribution of multi-scale features;

[0204] S514, Path Generation and Evaluation:

[0205] Path generation: The generator inputs random noise, environmental characteristics and path constraints to generate potential paths. The calculation formula is:

[0206] ;

[0207] Path evaluation: The discriminator inputs the generated path, environmental features, and multi-scale features, and outputs the feasibility score of the path. The calculation formula is:

[0208] ;

[0209] True path evaluation: The discriminator inputs the true path, environmental features, and multi-scale features, and outputs the feasibility score of the true path. The calculation formula is:

[0210] ;

[0211] Through the above content, combined with environmental characteristics, path constraints and multi-scale features, the generator can generate potential paths that are more adapted to the current environment and navigation needs, while improving the accuracy of the discriminator in evaluating the feasibility of the path, thereby achieving more intelligent and efficient navigation decisions.

[0212] The path selection in S52 includes:

[0213] S521, path scoring sorting: sort all generated potential paths Feasibility score Sort to get a list of paths sorted in descending order by score , corresponding to , the calculation formula is:

[0214] ;

[0215] in, For the discriminator path The feasibility score, is the environmental feature vector, It is a multi-scale feature;

[0216] S522, select the best path: select the path with the highest feasibility score from the sorted path list As the current navigation path, the calculation formula is:

[0217] ;

[0218] in, The path with the highest feasibility score is Give the corresponding score;

[0219] Through the above content, the potential paths generated by the generative adversarial network are sorted and screened to ensure that the path with the highest feasibility score is selected as the current navigation path, which improves the reliability and accuracy of navigation decisions and ensures that the robot can navigate efficiently and safely in complex environments.

[0220] Path correction in S53 includes:

[0221] S531, Environmental Change Monitoring: Use cameras and sensors to obtain the latest image data and sensor data of the robot's surroundings in real time, and update the environmental feature vector , the calculation formula is:

[0222] ;

[0223] in, and are the latest image data and sensor data, respectively. Extract function for environmental features;

[0224] S532, Path Regeneration: Generator Input Random Noise , the updated environmental feature vector and path constraints , generating new potential paths , the calculation formula is:

[0225] ;

[0226] in, is the generator function, is the weight matrix of the generator, is the first bias, is the activation function;

[0227] S533, Path re-evaluation: The discriminator inputs a new potential path , the updated environmental feature vector and multi-scale features , output the new path feasibility score , the calculation formula is:

[0228] ;

[0229] in, is the discriminator function, is the weight matrix of the discriminator, is the second bias, is the activation function;

[0230] S534, Path selection and switching: score the new path based on its feasibility , choose the path with the highest feasibility score As the current navigation path, the robot navigation path is switched in real time. The calculation formula is:

[0231] ;

[0232] Through the above content, combined with real-time monitoring of environmental changes, the generative adversarial network is used to dynamically generate and evaluate new potential paths, and the path with the highest feasibility score is selected for navigation, ensuring that the robot can flexibly respond to complex and dynamic environments and improve the safety and reliability of navigation.

[0233] Navigation decisions and dynamic adjustments in S6 include:

[0234] S61, Navigation decision: Generate navigation instructions based on the results of path prediction and correction, combined with the robot's current position, speed and direction, to guide the robot to perform real-time navigation, including:

[0235] S611, path information acquisition: obtain the currently selected best path from the path correction ;

[0236] S612, status information update: obtain the current position of the robot in real time ,speed and direction , the calculation formula is:

[0237] ;

[0238] S613, generating navigation instructions: according to the best path and the current state of the robot, and generates navigation instructions using the model predictive control (MPC) algorithm , guiding the robot to perform real-time navigation, the calculation formula is:

[0239] ;

[0240] in, For control inputs (such as speed and steering angle), and is the weight matrix, is the Euclidean distance;

[0241] S62, Dynamic Adjustment: During robot navigation, the robot monitors environmental changes and robot status changes in real time, and dynamically adjusts navigation decisions to ensure that the robot can adapt to environmental changes, avoid obstacles, and successfully reach the target point. Specifically, it includes:

[0242] Environmental change monitoring: Use cameras and sensors to monitor environmental changes in real time and update environmental feature vectors ,

[0243] Status feedback update: Update the robot's status information in real time , including position, speed and direction, the calculation formula is:

[0244] ;

[0245] Navigation instruction adjustment: adjust navigation instructions according to the latest environmental characteristics and status feedback , to ensure that the robot can dynamically avoid obstacles and navigate smoothly, the calculation formula is:

[0246] ;

[0247] Through the above steps, combined with the results of path prediction and correction, the robot's current position, speed and direction, the navigation instructions are generated and dynamically adjusted to achieve efficient and real-time navigation decisions. The model predictive control algorithm is used to generate the optimal navigation instructions, and the environment and state changes are monitored in real time during the navigation process. The navigation decisions are dynamically adjusted to ensure that the robot can flexibly avoid obstacles and reach the target point smoothly. The accuracy, reliability and adaptability of the navigation system are significantly improved, and the robot's autonomous navigation capability in complex environments is enhanced.

[0248] Feedback mechanisms and optimizations in S7 include:

[0249] S71, Feedback Mechanism: Establish a feedback mechanism through real-time image data and navigation results to continuously optimize target recognition and navigation decisions, including:

[0250] S711, Real-time Data Acquisition: Use cameras and sensors to acquire image data and sensor data in real time during robot navigation;

[0251] S712, Navigation result recording: Record the path selection, obstacle avoidance and navigation effect of the robot during navigation, including the actual path , obstacle avoidance points and time to reach the target ;

[0252] S713, data feedback: Feedback the real-time acquired image data, sensor data and navigation results. The calculation formula is:

[0253] ;

[0254] in, is real-time image data, is real-time sensor data, is the navigation result;

[0255] S72, target recognition optimization: Based on the results of the feedback mechanism, the weights of the convolutional neural network (CNN) model are updated using an online learning algorithm. The calculation formula is:

[0256] ;

[0257] in, are model parameters, is the learning rate, is the loss function, is the gradient calculation;

[0258] S73, Navigation decision optimization: According to the results of the feedback mechanism, the navigation strategy is updated using the reinforcement learning algorithm. The calculation formula is:

[0259] ;

[0260] in, is the current state, is the action, s' is the state at the next moment, a' is the action at the next moment, For instant rewards, is the discount factor, is the learning rate;

[0261] Through the above content, combined with real-time image data and navigation result feedback, the target recognition and navigation decision-making system is dynamically updated and optimized to ensure that the robot always maintains efficient and safe navigation capabilities in a constantly changing environment, thereby improving the overall performance and reliability of the system.

[0262] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention is limited to these examples. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity.

Claims

1. A robot navigation decision method based on target recognition, characterized in that: The following steps are involved: S1: Use cameras and image sensors to obtain real-time image data of the robot's surrounding environment, and pre-process the acquired image data, including image denoising, grayscale conversion, and edge detection; S2: Perform target recognition on the preprocessed image data, identify the target object in the environment, and extract feature information from the identified target object; S3: Integrate feature information and use a grid map to construct the integrated feature information into a grid environment model to generate an environment map, which includes the location, shape and related information of the target object; S4: Based on the constructed environment model, calculate the optimal path from the current position of the robot to the target point; S5: Based on the results of path planning, a generative adversarial network is introduced to predict and correct the path, dynamically adjust and optimize the robot's navigation path, including: S51: Generate potential paths using a generative adversarial network, evaluate the generated potential paths, and output the feasibility score of each generated path; specifically, the generator not only inputs random noise, but also inputs the current environment model, robot position information and path constraints to generate potential paths; the discriminator inputs the generated path and the real path, and evaluates them in combination with environmental features and multi-scale features, and outputs the feasibility score of the path; through adversarial training of the generator and the discriminator; the loss function of the generator is: ； in, is random noise, is the environmental feature vector, is the path constraint, is a multi-scale feature, is the probability distribution of noise, is the probability distribution of environmental characteristics, is the probability distribution of path constraints, D represents the identifier, It is expectation; S52: According to the feasibility score result, select the path with the highest feasibility as the current navigation path; S53: During the actual navigation process of the robot, the environment changes are monitored in real time, and the path is dynamically corrected through the generative adversarial network; S6: Generate navigation instructions based on the results of path prediction and correction, combined with the robot's current position, speed and direction, to guide the robot to perform real-time navigation and dynamically adjust navigation decisions; S7: Continuously optimize target recognition and navigation decisions through real-time image data and navigation result feedback.

2. The robot navigation decision method based on target recognition according to claim 1 is characterized in that: The S1 step includes: S11, image data acquisition: using a camera and an image sensor installed on the robot to acquire real-time image data of the robot's surroundings; S12, image denoising: using a Gaussian filter to denoise the acquired image data; S13, image grayscale processing: grayscale the denoised image data using a weighted average method to convert the color image into a grayscale image; S14, edge detection processing: using the Canny edge detection algorithm to perform edge detection on the grayscale image to extract edge information in the image.

3. The robot navigation decision method based on target recognition according to claim 1 is characterized in that: The S2 step includes: S21, target recognition: Use the convolutional neural network model to perform target recognition on the preprocessed image data to identify the target objects in the environment, including obstacles, target points and navigation signs; S22, feature extraction: extracting feature information from the identified target object, including the shape, size, position and motion state of the target object, specifically including: S221, shape feature extraction: using the Canny edge detection algorithm to extract the contour and edge of the target object, and calculate the shape features, including perimeter, area and contour approximation polygon; S222, size feature extraction: calculate the bounding box of the target object to obtain its width, height and area; S223, position feature extraction: determine the position of the target object in the image and calculate its centroid coordinates; S224, motion state feature extraction: calculate the motion trajectory and speed of the target object through continuous multi-frame image data, and estimate the motion vector of the target object using an optical flow algorithm.

4. The robot navigation decision method based on target recognition according to claim 3 is characterized in that: Step S3 specifically includes: S31, feature information integration: integrating the feature information of the target object obtained from target recognition and feature extraction; S32, environmental model construction: Based on the integrated feature information, a grid environment model is constructed using a grid map method; the grid map divides the environment into a number of grids, each grid contains a probability value, indicating the possibility of the corresponding area being occupied, and the occupancy probability of the grid is updated using the position and size information of the target object to generate a grid map containing the target object; the mathematical formula of the grid environment model is: ； in, It is The probability that a grid is occupied is is the log-odds of the grid, From time 1 to The sensor measurement value of S33, environment map generation: converting the constructed environment model into an environment map.

5. The robot navigation decision method based on target recognition according to claim 4 is characterized in that: The S4 step includes: S41, path planning: Based on the constructed grid environment model, the A* algorithm is used to calculate the optimal path from the current position of the robot to the target point; S42, path optimization: In the path planning process, the robot size and motion capability are considered, and the path is smoothed using the spline interpolation method. The calculation formula is: ； in, is a point on the smooth path, is the control point, as a parameter.

6. The robot navigation decision method based on target recognition according to claim 1 is characterized in that: The step S51 includes: S511, the generator calculation formula is: ； in, is random noise, is the environmental feature vector, is the path constraint, is the weight matrix of the generator, is the first bias, is the activation function, Represents a generator; S512, the discriminator calculation formula is: ； in, is the input path, is the environmental feature vector, is a multi-scale feature, is the weight matrix of the discriminator, is the second bias, is the activation function; S513, in generative adversarial training, the loss function of the discriminator is: ； in, is the real path data, is the environmental feature vector, is a multi-scale feature, is the probability distribution of the real path data, is the probability distribution of environmental characteristics, is the probability distribution of multi-scale features; S514, Path Generation and Evaluation: Path generation: The generator inputs random noise, environmental characteristics and path constraints to generate potential paths. The calculation formula is: ； Path evaluation: The discriminator inputs the generated path, environmental features, and multi-scale features, and outputs the feasibility score of the path. The calculation formula is: ； True path evaluation: The discriminator inputs the true path, environmental features, and multi-scale features, and outputs the feasibility score of the true path. The calculation formula is: 。 7. The robot navigation decision method based on target recognition according to claim 6 is characterized in that: The step S52 includes: S521, path scoring sorting: sort all generated potential paths Feasibility score Sort to get a list of paths sorted in descending order by score , corresponding to , the calculation formula is: ； in, For the discriminator path The feasibility score, is the environmental feature vector, It is a multi-scale feature; S522, select the best path: select the path with the highest feasibility score from the sorted path list As the current navigation path, the calculation formula is: ； in, The path with the highest feasibility score is Give it the corresponding rating.

8. The robot navigation decision method based on target recognition according to claim 7 is characterized in that: The step S53 includes: S531, Environmental Change Monitoring: Use cameras and sensors to obtain the latest image data and sensor data of the robot's surroundings in real time, and update the environmental feature vector , the calculation formula is: ； in, and are the latest image data and sensor data, respectively. Extract function for environmental features; S532, Path Regeneration: Generator Input Random Noise , the updated environmental feature vector and path constraints , generating new potential paths , the calculation formula is: ； in, is the generator function, is the weight matrix of the generator, is the first bias, is the activation function; S533, Path re-evaluation: The discriminator inputs a new potential path , the updated environmental feature vector and multi-scale features , output the new path feasibility score , the calculation formula is: ； in, is the discriminator function, is the weight matrix of the discriminator, is the second bias, is the activation function; S534, Path selection and switching: score the new path based on its feasibility , choose the path with the highest feasibility score As the current navigation path, the robot navigation path is switched in real time. The calculation formula is: 。 9. The robot navigation decision method based on target recognition according to claim 8, characterized in that: The S6 step includes: S61, Navigation decision: Generate navigation instructions based on the results of path prediction and correction, combined with the robot's current position, speed and direction, to guide the robot to perform real-time navigation, including: S611, path information acquisition: obtain the currently selected best path from the path correction ; S612, status information update: obtain the current position of the robot in real time ,speed and direction , the calculation formula is: ； S613, generating navigation instructions: according to the best path and the current state of the robot, and generates navigation instructions using a model predictive control algorithm , guiding the robot to perform real-time navigation, the calculation formula is: ； in, is the control input, and is the weight matrix, is the Euclidean distance, represents the moment, T is the total navigation time; S62, dynamic adjustment: During the robot navigation process, the environment changes and robot state changes are monitored in real time, and navigation decisions are adjusted dynamically.

10. The robot navigation decision method based on target recognition according to claim 9, characterized in that: The S7 step includes: S71, Feedback Mechanism: Establish a feedback mechanism through real-time image data and navigation results to continuously optimize target recognition and navigation decisions, including: S711, Real-time Data Acquisition: Use cameras and sensors to acquire image data and sensor data in real time during robot navigation; S712, Navigation result recording: Record the path selection, obstacle avoidance and navigation effect of the robot during navigation, including the actual path , obstacle avoidance points and time to reach the target ; S713, data feedback: Feedback the real-time acquired image data, sensor data and navigation results. The calculation formula is: ； in, is real-time image data, is real-time sensor data, is the navigation result; S72, target recognition optimization: According to the results of the feedback mechanism, the weights of the convolutional neural network model are updated using an online learning algorithm. The calculation formula is: ； in, are model parameters, is the learning rate, is the loss function, is the gradient calculation; S73, Navigation decision optimization: According to the results of the feedback mechanism, the navigation strategy is updated using the reinforcement learning algorithm. The calculation formula is: ； in, is the current state, is the action, s' is the state at the next moment, a' is the action at the next moment, For instant rewards, is the discount factor, is the learning rate.

Images