Robot navigation and obstacle avoidance method and system based on meta-learning

Abstract

The present invention relates to the technical field of navigation and obstacle avoidance, and discloses a robot navigation and obstacle avoidance method and system based on meta-learning. The method comprises: acquiring an image within the current field of view of the robot; inputting the acquired image into a trained navigation and obstacle avoidance model, and outputting identified obstacle areas and obstacle types; acquiring current basic information of the robot; processing the current basic information, obstacle areas and obstacle types of the robot based on a path planning algorithm to obtain an obstacle avoidance path for the robot; the working principle of the trained navigation and obstacle avoidance model comprises: extracting features from the acquired image to obtain a feature map; generating multiple bounding boxes based on the feature map; determining the position of the obstacle based on the bounding boxes; correcting the size of each bounding box to determine the area of ​​the obstacle; converting features within bounding box areas of different sizes into a feature map of fixed size; and identifying the type of the obstacle based on the feature map of fixed size.

Description

Technical Field

[0001] The present invention relates to the field of navigation and obstacle avoidance technology, and in particular to a robot navigation and obstacle avoidance method and system based on meta-learning. Background Art

[0002] The statements in this section merely mention background art related to the present invention and do not necessarily constitute prior art.

[0003] Intelligent robotics technologies are developing rapidly. Robots capable of replacing humans in high-risk and repetitive tasks are a hot topic for researchers. Mobility is essential for robots, and the core technologies of robotic navigation systems are positioning, path planning, and sensor systems. However, path planning systems present challenges when obstacles occur during robot movement, making it difficult to adjust the robot's posture or effectively plan a path. This is especially true when the robot is in confined areas or inaccessible obstacles, making it difficult to ensure operational feasibility, safety, and rationality.

[0004] With the continuous advancement of image processing technology, machine vision-based image recognition has achieved significant breakthroughs. Deep learning-based image recognition technology is poised to replace manually designed feature extraction algorithms, improving efficiency and conserving resources. However, applying deep learning algorithms to identify obstacles in robot navigation presents a challenge: obstacles come in a wide variety of styles, ranging from traversable to impassable, and from dangerous to safe. Summary of the Invention

[0005] In order to address the deficiencies of the prior art, the present invention provides a robot navigation and obstacle avoidance method and system based on meta-learning; the present invention can realize dynamic obstacle avoidance of the robot.

[0006] In a first aspect, the present invention provides a robot navigation and obstacle avoidance method based on meta-learning;

[0007] Robot navigation and obstacle avoidance methods based on meta-learning, including:

[0008] Get the image within the robot's current field of view;

[0009] Input the acquired image into the trained navigation and obstacle avoidance model, and output the identified obstacle area and obstacle type;

[0010] Obtain the robot's current basic information; based on the path planning algorithm, process the robot's current basic information, obstacle area, and obstacle type to obtain the robot's obstacle avoidance path; the basic information includes: the robot's current position coordinates, target point coordinates, its own posture, and speed information;

[0011] The working principle of the trained navigation and obstacle avoidance model includes: extracting features from the acquired image to obtain a feature map; generating multiple bounding boxes based on the feature map; determining the location of the obstacle based on the bounding boxes; correcting the size of each bounding box to determine the area of ​​the obstacle; converting the features within the bounding box areas of different sizes into a feature map of a fixed size; and identifying the type of obstacle based on the feature map of the fixed size.

[0012] In a second aspect, the present invention provides a robot navigation and obstacle avoidance system based on meta-learning;

[0013] The robot navigation and obstacle avoidance system based on meta-learning includes:

[0014] An acquisition module is configured to: acquire an image within the current field of view of the robot;

[0015] A recognition module is configured to: input the acquired image into the trained navigation obstacle avoidance model and output the identified obstacle area and obstacle type;

[0016] A path planning module is configured to: obtain the current basic information of the robot; process the current basic information, obstacle area, and obstacle type of the robot based on the path planning algorithm to obtain the robot's obstacle avoidance path; the basic information includes: the robot's current position coordinates, target point coordinates, its own posture, and speed information;

[0017] The working principle of the trained navigation and obstacle avoidance model includes: extracting features from the acquired image to obtain a feature map; generating multiple bounding boxes based on the feature map; determining the location of the obstacle based on the bounding boxes; correcting the size of each bounding box to determine the area of ​​the obstacle; converting the features within the bounding box areas of different sizes into a feature map of a fixed size; and identifying the type of obstacle based on the feature map of the fixed size.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] This meta-learning-based robot navigation and obstacle avoidance method, compared to previous methods, utilizes metric meta-learning to construct different task datasets from small sample data. This method, on the one hand, addresses deep learning's reliance on massive datasets, thereby enabling deep learning algorithms to be trained with only a small number of samples. Furthermore, given the diverse types of obstacles, identifying them can affect the model's accuracy and speed. Therefore, to adapt the algorithm to different and unknown types of obstacles, this method introduces meta-learning to construct task-level datasets, enhancing the model's generalization performance and enabling the model to learn correlations between different types, thus enabling the algorithm to handle a wide range of obstacle recognition challenges. This method is ultimately integrated with the Transient Object-Engineered (TEB) algorithm to create a safe and efficient dynamic obstacle avoidance system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0021] Figure 1 A schematic diagram of a network learning process for roadblock recognition based on meta-learning provided by the present invention;

[0022] Figure 2 This is a schematic diagram of the network structure of the first embodiment of the present invention;

[0023] Figure 3 This is a detailed diagram of the network structure of Example 1 of the present invention. DETAILED DESCRIPTION

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0025] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0027] All data in this embodiment is obtained in compliance with laws and regulations and based on the consent of the user, and is used legally.

[0028] Timed-Elastic-Band (TEB) is a method for calculating collision-free trajectories around obstacles in the field of robotic navigation. It uses real-time sensor information to identify obstacle areas and designs routes based on the principle of minimum curvature.

[0029] Example 1

[0030] This embodiment provides a robot navigation and obstacle avoidance method based on meta-learning;

[0031] like Figure 1 As shown in the figure, the robot navigation and obstacle avoidance method based on meta-learning includes:

[0032] S101: Acquire the image within the robot's current field of view;

[0033] S102: Input the acquired image into the trained navigation and obstacle avoidance model, and output the identified obstacle area and obstacle type;

[0034] S103: Obtain the current basic information of the robot; based on the path planning algorithm, process the current basic information of the robot, obstacle area and obstacle type to obtain the obstacle avoidance path of the robot; the basic information includes: the current position coordinates of the robot, the coordinates of the target point, its own posture and speed information;

[0035] The working principle of the trained navigation and obstacle avoidance model includes: extracting features from the acquired image to obtain a feature map; generating multiple bounding boxes based on the feature map; determining the location of the obstacle based on the bounding boxes; correcting the size of each bounding box to determine the area of ​​the obstacle; converting the features within the bounding box areas of different sizes into a feature map of a fixed size; and identifying the type of obstacle based on the feature map of the fixed size.

[0036] Furthermore, if Figure 2 As shown, the trained navigation obstacle avoidance model has a network structure including:

[0037] The feature vector extraction module, target position detection module, target feature extraction module and target recognition module are connected in sequence.

[0038] Furthermore, if Figure 3 As shown, the feature vector extraction module includes: a first convolutional layer, a second convolutional layer, a first maximum pooling layer, a third convolutional layer, a fourth convolutional layer and a second maximum pooling layer connected in sequence; wherein, the connection end between the second maximum pooling layer and the fourth convolutional layer is the input end of the second maximum pooling layer; and the non-connected end between the second maximum pooling layer and the fourth convolutional layer is the output end of the second maximum pooling layer.

[0039] Furthermore, the feature vector extraction module works as follows:

[0040] Perform feature extraction on the acquired image to obtain a feature map.

[0041] During the model training phase, the images to be identified for different tasks are input into the feature vector extraction module to extract the common features of the images.

[0042] In actual use, you only need to input an image to be tested to obtain the feature map.

[0043] By mapping samples of different classes into the same feature space and measuring the distances between samples of different types, we can further learn the shared feature maps between obstacle samples of different types. Different types of obstacles also share common characteristics. For example, obstacles modeled on humans and obstacles modeled on animals, while different in type, are both living things. Only by extracting high-dimensional features can the network recognize new, unseen types.

[0044] Furthermore, the target location detection module includes: a candidate region network (RPN: Region Proposal Networks);

[0045] The input end of the candidate region network is connected to the output end of the second maximum pooling layer;

[0046] The output end of the candidate region network block is connected to the input end of the fifth convolutional layer;

[0047] The output of the fifth convolutional layer is connected to the input of the sixth convolutional layer;

[0048] The output of the sixth convolutional layer is connected to the input of the seventh convolutional layer.

[0049] Furthermore, the target position detection module operates as follows:

[0050] Based on the feature map, multiple bounding boxes are generated; based on the bounding boxes, the location of the obstacle is determined; the size of each bounding box is corrected to determine the area of ​​the obstacle.

[0051] First, multiple fixed-size bounding boxes are generated based on the obtained feature map. The scale and aspect ratio of the anchor boxes are set, and the anchor boxes are slid on the feature map with a set step size to obtain the corresponding bounding boxes. The number of bounding boxes is set to K.

[0052] Then, the corresponding features of the obtained set of bounding boxes are input into the convolution layer of 3×3 convolution kernel to obtain the processed features.

[0053] The processed features are sequentially fed into two convolutional layers with 1×1 convolution kernels. One convolutional layer has 2K channels, corresponding to the two predicted values ​​for each bounding box. Classification is then performed using a normalized exponential function to indicate the probability of whether a bounding box contains an object. Finally, the classification convolutional layer is constrained by the binary cross entropy loss function. The other convolutional layer has 4K channels, corresponding to the offsets in the four directions of each bounding box. The regression convolutional layer is constrained by the smooth L1 loss.

[0054] Binary Cross Entropy Loss:

[0055]

[0056] in, Indicates the The bounding box predicts the probability value of containing the target, Indicates the The true probability value of the bounding box containing the target, Indicates the batch size.

[0057] L1 loss function Smooth L1 Loss:

[0058]

[0059] in, Indicates the Four offsets for bounding box predictions, Indicates the The actual offset of the bounding box, Indicates the size of the offset, Indicates the number of bounding boxes.

[0060] Finally, the bounding box is corrected using the predicted four offsets.

[0061] Furthermore, the candidate region network is used to generate multiple bounding boxes based on the feature map.

[0062] Furthermore, the sixth convolutional layer is used to determine the position of the obstacle based on the bounding box.

[0063] Furthermore, the seventh convolutional layer is used to correct the size of each bounding box to determine the area of ​​the obstacle.

[0064] The extracted features are input into the candidate region network to generate multiple candidate regions containing the target, and classification loss and regression loss are generated to constrain the candidate region network.

[0065] Furthermore, the target feature extraction module is a third maximum pooling layer;

[0066] An input end of the third maximum pooling layer is connected to an output end of the seventh convolutional layer.

[0067] The feature map is further extracted from the target area pooling to extract the fixed-size features of the region of interest;

[0068] Furthermore, the target feature extraction module works as follows:

[0069] The features within the bounding box regions of different sizes are converted into feature maps of fixed size.

[0070] In order to cope with the feature maps corresponding to bounding boxes of different sizes, a maximum pooling layer is introduced. The maximum pooling layer is used to convert the features in the bounding box area of ​​interest into a feature map of fixed scale, which facilitates subsequent target recognition.

[0071] Furthermore, the target recognition module includes: a recurrent neural network;

[0072] The input end of the recurrent neural network is connected to the output end of the third maximum pooling layer;

[0073] The output end of the recurrent neural network is used to output the recognition result.

[0074] Furthermore, the target recognition module operates as follows:

[0075] For a fixed-size feature map, identify the type of obstacle.

[0076] The obtained features are input into the recurrent neural network to obtain the recognition results and regression results. The generated classification loss and regression loss are used to constrain the recurrent neural network.

[0077] After obtaining the fixed-size feature map, it is input into two different fully connected layers, one of which has neurons, of which is the total number of recognized obstacle categories. After this fully connected layer, a normalized exponential function is added to calculate the probability value of the recognized obstacle, and the classification network module is constrained by the cross entropy loss (Cross Entropy Loss); another fully connected layer has neurons, used to further regress the offsets of the bounding box in four directions, and constrain the regression network module through the smooth L1 loss Smooth L1 Loss.

[0078] Cross Entropy Loss:

[0079] ;

[0080] in, Indicates the type of obstacle to be identified. Indicates identification The probability value of the class.

[0081] Smooth L1 Loss:

[0082]

[0083] in, Indicates that the bounding box content is recognized as The regression offset of the class, Indicates the true offset of the bounding box, Indicates the offset size, Indicates that the identified content is background and does not participate in loss calculation.

[0084] Furthermore, the training process of the trained navigation obstacle avoidance model includes:

[0085] Constructing a training set and a test set; both the training set and the test set are images of known obstacle areas and obstacle types;

[0086] Input the training set into the navigation and obstacle avoidance model and train the model. When the total loss function value no longer decreases or the number of iterations exceeds the set number, stop training and obtain a preliminary navigation and obstacle avoidance model.

[0087] The preliminary navigation and obstacle avoidance model is tested using the test set. When the test accuracy exceeds the set threshold, the test is stopped and the current navigation and obstacle avoidance model is the final trained navigation and obstacle avoidance model. If the test accuracy is lower than the set threshold, the training set is replaced and training is performed again.

[0088] Furthermore, the total loss function is equal to the weighted sum of the first classification loss function, the first regression loss function, the second classification loss function and the second regression loss function.

[0089] Among them, the first classification loss function and the first regression loss function are the loss functions of the sixth convolutional layer and the seventh convolutional layer;

[0090] Among them, the second classification loss function and the second regression loss function are loss functions of the recurrent neural network.

[0091] The total loss function (Loss) is calculated by adding the weighted sum of Binary Cross Entropy Loss, Cross Entropy Loss, and two Smooth L1 Losses. The network parameters are updated using the backpropagation algorithm. The network is trained until the number of training rounds reaches the set value or the loss shows a stable trend. The network model with the minimum loss value is saved.

[0092] Furthermore, the total loss function is expressed as:

[0093]

[0094] in, and is a hyperparameter used to balance various loss functions.

[0095] Furthermore, the training and test sets are constructed; each obstacle classification is considered a task, and different tasks are constructed for different types of obstacle images. Different obstacle classifications require different tasks, such as identifying living obstacles, identifying inanimate obstacles, and identifying static obstacles. Each task constructs a dataset specifically for its own domain and performs identification.

[0096] Furthermore, the construction of the training set and the test set specifically includes:

[0097] First, we group all types of roadblock images by category, treating each category as a separate recognition task. All tasks form a complete dataset, which we then divide into training, validation, and test sets. The validation and test sets contain the same image types.

[0098] During the training process, C types of tasks are randomly selected from the training set in each round, and K samples are selected from each type of task to form a sample set:

[0099] .

[0100] Select a batch of samples from the remaining samples of this C-type task to form the image set to be identified The construction of validation set and test set is similar.

[0101] The training set is used to train the model, the validation set is used to test the performance of the model during training, and the test set is used to test the performance of the model after training.

[0102] Furthermore, the path planning algorithm is implemented using a Timed-Elastic-Band (TEB) algorithm.

[0103] This invention achieves better recognition of both known and unknown obstacle images by modeling multiple types of obstacle image data. This process can be considered an example of meta-learning, as the goal of learning is to better solve the task (i.e., achieve stronger generalization performance to unknown samples), which is much more difficult than simply learning a model with fixed data. To reduce the requirement for large amounts of training data and enhance the algorithm's versatility and practicality, this invention incorporates the concept of small-sample learning, achieving satisfactory performance using only a small number of training samples. The identified obstacle regions are then input into the TEB algorithm to calculate the optimal path around the obstacles, ensuring the safety and reliability of robot operation. Specifically, this invention uses a metric approach to construct different tasks for different types of obstacle images. These task images are then mapped onto the same feature space to generate feature maps. A Region Proposal Network (RPN) is then used to locate targets within the feature maps. Region of Interest (ROI) pooling is then used to extract target features. Finally, a Recurrent Neural Network (R-CNN) is used to classify and identify targets. The identified target regions and their types are then input into the TEB algorithm for dynamic obstacle avoidance route planning. Based on the concept of meta-learning, each type of roadblock recognition is considered a task. To reduce the need for massive data training, this paper employs the concept of small-sample learning, training on only a small number of samples from each task. The training results are then applied to new roadblock recognition tasks, improving the model's practicality.

[0104] The image obtained by the robot camera is input into the recognition network to identify the roadblock area and type Finally, the sensor and positioning system are used to obtain the current robot's position coordinates, target point coordinates, its own posture, speed and other information Input TEB algorithm to get the optimal obstacle avoidance path .

[0105] .

[0106] Example 2

[0107] This embodiment provides a robot navigation and obstacle avoidance system based on meta-learning;

[0108] The robot navigation and obstacle avoidance system based on meta-learning includes:

[0109] An acquisition module is configured to: acquire an image within the current field of view of the robot;

[0110] A recognition module is configured to: input the acquired image into the trained navigation obstacle avoidance model and output the identified obstacle area and obstacle type;

[0111] A path planning module is configured to: obtain the current basic information of the robot; process the current basic information, obstacle area, and obstacle type of the robot based on the path planning algorithm to obtain the robot's obstacle avoidance path; the basic information includes: the robot's current position coordinates, target point coordinates, its own posture, and speed information;

[0112] The working principle of the trained navigation and obstacle avoidance model includes: extracting features from the acquired image to obtain a feature map; generating multiple bounding boxes based on the feature map; determining the location of the obstacle based on the bounding boxes; correcting the size of each bounding box to determine the area of ​​the obstacle; converting the features within the bounding box areas of different sizes into a feature map of a fixed size; and identifying the type of obstacle based on the feature map of the fixed size.

[0113] It should be noted here that the above-mentioned acquisition module, identification module and path planning module correspond to steps S101 to S103 in Example 1. The examples and application scenarios implemented by the above-mentioned modules and corresponding steps are the same, but are not limited to the contents disclosed in the above-mentioned Example 1.

[0114] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A robot navigation and obstacle avoidance method based on meta-learning, characterized by: include: Get the image within the robot's current field of view; Input the acquired image into the trained navigation and obstacle avoidance model, and output the identified obstacle area and obstacle type; Obtaining the current basic information of the robot; processing the current basic information, obstacle area, and obstacle type of the robot based on a path planning algorithm to obtain an obstacle avoidance path for the robot, wherein the path planning algorithm is implemented using a time elastic band algorithm; the basic information includes: the robot's current position coordinates, target point coordinates, its own posture, and speed information; The working principle of the trained navigation and obstacle avoidance model includes: extracting features from the acquired image to obtain a feature map; generating multiple bounding boxes based on the feature map; determining the location of the obstacle based on the bounding boxes; correcting the size of each bounding box to determine the area of ​​the obstacle; converting the features within the bounding box areas of different sizes into a feature map of a fixed size; and identifying the type of obstacle based on the feature map of the fixed size. The trained navigation and obstacle avoidance model has a network structure comprising: The feature vector extraction module, target position detection module, target feature extraction module and target recognition module are connected in sequence, wherein the target position detection module includes: a candidate area network; the target feature extraction module is the third maximum pooling layer; the target recognition module includes: a recurrent neural network.

2. The robot navigation and obstacle avoidance method based on meta-learning as claimed in claim 1, characterized in that: The feature vector extraction module includes: a first convolutional layer, a second convolutional layer, a first maximum pooling layer, a third convolutional layer, a fourth convolutional layer, and a second maximum pooling layer connected in sequence; wherein the connection end between the second maximum pooling layer and the fourth convolutional layer is the input end of the second maximum pooling layer; and the non-connected end between the second maximum pooling layer and the fourth convolutional layer is the output end of the second maximum pooling layer.

3. The robot navigation and obstacle avoidance method based on meta-learning as claimed in claim 1, characterized in that: The input end of the candidate region network is connected to the output end of the second maximum pooling layer; The output end of the candidate region network block is connected to the input end of the fifth convolutional layer; The output of the fifth convolutional layer is connected to the input of the sixth convolutional layer; The output of the sixth convolutional layer is connected to the input of the seventh convolutional layer.

4. The robot navigation and obstacle avoidance method based on meta-learning as claimed in claim 1, characterized in that: The target position detection module works as follows: First, multiple fixed-size bounding boxes are generated based on the obtained feature map. The scale and aspect ratio of the anchor box are set, and the anchor box is slid on the feature map according to the set step size to obtain the corresponding bounding box. Then, the corresponding features of the obtained set of bounding boxes are input into the convolution layer of the 3×3 convolution kernel to obtain the processed features; The processed features are sequentially input into two convolutional layers with 1×1 convolution kernels. One of the convolutional layers has 2K channels, corresponding to the two predicted values ​​of each bounding box. The normalized exponential function is then used for classification to indicate the probability of whether the bounding box contains the target. Finally, the classification convolutional layer is constrained by the binary cross entropy loss function. The other convolutional layer has 4K channels, corresponding to the offsets in the four directions of each bounding box, and the regression convolutional layer is constrained by the smooth L1 loss.

5. The robot navigation and obstacle avoidance method based on meta-learning as claimed in claim 1, characterized in that: An input end of the third maximum pooling layer is connected to an output end of the seventh convolutional layer.

6. The robot navigation and obstacle avoidance method based on meta-learning as claimed in claim 1, characterized in that: The input end of the recurrent neural network is connected to the output end of the third maximum pooling layer; the output end of the recurrent neural network is used to output the recognition result.

7. The robot navigation and obstacle avoidance method based on meta-learning as claimed in claim 1, characterized in that: The target recognition module works as follows: The obtained features are input into the recurrent neural network to obtain the recognition results and regression results. The generated classification loss and regression loss are used to constrain the recurrent neural network; After obtaining the fixed-size feature map, it is input into two different fully connected layers, one of which has neurons, of which is the total number of recognized obstacle categories; after this fully connected layer, a normalized exponential function is added to calculate the probability value of the recognized obstacle, and the classification network module is constrained by the cross entropy loss; another fully connected layer has neurons, used to further regress the offsets of the bounding box in four directions, and constrain the regression network module through the smooth L1 loss SmoothL1 Loss.

8. The robot navigation and obstacle avoidance method based on meta-learning as claimed in claim 3, characterized in that: The training process of the trained navigation obstacle avoidance model includes: Constructing a training set and a test set; both the training set and the test set are images of known obstacle areas and obstacle types; Input the training set into the navigation and obstacle avoidance model and train the model. When the total loss function value no longer decreases or the number of iterations exceeds the set number, stop training and obtain a preliminary navigation and obstacle avoidance model. The preliminary navigation and obstacle avoidance model is tested using the test set. When the test accuracy exceeds the set threshold, the test is stopped and the current navigation and obstacle avoidance model becomes the final trained navigation and obstacle avoidance model. If the test accuracy is lower than the set threshold, the training set is replaced and training is repeated. The total loss function is equal to the weighted sum of the first classification loss function, the first regression loss function, the second classification loss function and the second regression loss function; wherein the first classification loss function and the first regression loss function are the loss functions of the sixth convolutional layer and the seventh convolutional layer; wherein the second classification loss function and the second regression loss function are the loss functions of the recurrent neural network.

9. A robot navigation and obstacle avoidance system based on meta-learning, characterized by: include: An acquisition module is configured to: acquire an image within the current field of view of the robot; A recognition module is configured to: input the acquired image into the trained navigation obstacle avoidance model and output the identified obstacle area and obstacle type; A path planning module is configured to: obtain the current basic information of the robot; process the current basic information, obstacle area, and obstacle type of the robot based on a path planning algorithm to obtain an obstacle avoidance path for the robot. The path planning algorithm is implemented using a time elastic band algorithm; the basic information includes the robot's current position coordinates, target point coordinates, its own posture, and speed information; The working principle of the trained navigation and obstacle avoidance model includes: extracting features from the acquired image to obtain a feature map; generating multiple bounding boxes based on the feature map; determining the location of the obstacle based on the bounding boxes; correcting the size of each bounding box to determine the area of ​​the obstacle; converting the features within the bounding box areas of different sizes into a feature map of a fixed size; and identifying the type of obstacle based on the feature map of the fixed size. The trained navigation and obstacle avoidance model has a network structure comprising: The feature vector extraction module, target position detection module, target feature extraction module and target recognition module are connected in sequence, wherein the target position detection module includes: a candidate area network; the target feature extraction module is the third maximum pooling layer; the target recognition module includes: a recurrent neural network.

Images