Method and system for using knowledge graph to assist navigation in intelligent navigation scenarios

Abstract

This invention provides a method and system for using knowledge graphs to assist navigation in intelligent navigation scenarios. First, a deep learning method is used to construct a multi-source data-fused navigation scenario knowledge graph, comgraph. A graph attention network (GAT) performs knowledge representation and knowledge reasoning on the knowledge graph, and meta-learning performs knowledge update and application in unknown scenarios to form navigation knowledge embedding, extracting spatial structural features. Simultaneously, a neural network is used to extract image features from the observed images in the scene, and semantic features are extracted using a target global word vector representation. The image features, semantic features, and spatial structural features are then spliced ​​into a multi-layer perceptron and a long short-term memory (MLP-LSTM) network for memory storage, and then transferred to reinforcement learning for state representation, enabling complete navigation. This invention enables intelligent agents to achieve excellent navigation results in new indoor navigation environments, regardless of changes in scene layout.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence robots, and in particular relates to a method for using knowledge graphs to assist navigation in intelligent navigation scenarios. Background Art

[0002] With the development of computer vision technology and reinforcement learning intelligent navigation, knowledge graph-assisted visual semantic navigation methods that are low-cost, highly efficient, independent of sensors (lidar, visual odometry, etc.), and adaptable to scene changes (strong lighting, weak texture, scene layout changes, etc.) have become current research hotspots. Knowledge graphs use computer vision technology to encode scene spatial information including target entities, relationships and attributes, and embed them into reinforcement learning navigation to improve the agent's spatial exploration and reasoning capabilities. Current navigation scene knowledge graphs are mostly derived from large-scale image semantic understanding datasets (VisualGenome), which are collected in real environments. Using this dataset in a virtual simulation environment has the problem of large target matching errors, and it is only used as a fixed spatial information feature as part of the reinforcement learning strategy, and cannot be adaptively updated in new scenes. Therefore, in response to the shortcomings of the existing technology, the present invention proposes a method for constructing a navigation knowledge graph and a method for effectively expressing and updating scene knowledge graphs using graph neural networks to improve the generalization ability of intelligent navigation. Summary of the Invention

[0003] In response to the problems of existing navigation knowledge graphs such as entity redundancy, target mismatch, inaccurate target relationships, and inability to accurately and effectively provide spatial information to assist navigation, the present invention proposes a method for constructing a common sense knowledge graph using deep learning methods in intelligent navigation scenarios, as well as a method for improving navigation generalization capabilities using graph neural networks and meta-learning update strategies.

[0004] In order to achieve the above-mentioned purpose, the technical solution of the present invention provides a method for using knowledge graphs to assist navigation in intelligent navigation scenarios. First, a deep learning method is used to construct a multi-source data fusion navigation scenario knowledge graph comgraph, and a graph attention network GAT for knowledge representation and knowledge reasoning of the knowledge graph and a meta-learning method for knowledge update and application of unknown scenarios are used to form navigation knowledge embedding to extract spatial structural features; at the same time, a neural network is used to extract image features of the observed images in the scene, and semantic features are extracted through the target global word vector representation; then the image features, semantic features and spatial structure features are spliced ​​into a multi-layer perceptron and a long short-term memory network MLP-LSTM for memory storage, and passed to reinforcement learning for state representation to perform complete navigation.

[0005] Moreover, based on the use of deep learning methods to construct a multi-source data fusion navigation scene knowledge graph, the navigation semantic knowledge is generalized to new scenes, so that the intelligent agent is not affected by changes in scene layout in the new indoor navigation environment.

[0006] Moreover, when the graph attention network GAT and meta-learning constitute navigation knowledge embedding, GAT embedding is used to represent and extract knowledge graph features as the input state of reinforcement learning navigation; meta-learning helps reinforcement learning to reduce the loss gradient during training and achieve generalization to new scenarios.

[0007] Moreover, the Resnet18 neural network is used to extract image features from the observed images in the scene.

[0008] Moreover, the image features, semantic features, and spatial structure features are spliced ​​into the MLP-LSTM neural network for memory storage and passed to reinforcement learning for state representation for complete navigation.

[0009] Moreover, based on the target position relationship in the navigation scene, the shortest path is inferred using prior knowledge, thereby giving reinforcement learning optimization decisions. The inference value is used as part of the action sampling to guide reinforcement learning navigation and improve the navigation generalization ability.

[0010] On the other hand, the present invention provides a system for using knowledge graphs to assist navigation in an intelligent navigation scenario, which is used to implement the method for using knowledge graphs to assist navigation in an intelligent navigation scenario as described above.

[0011] Moreover, it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the method of using knowledge graph to assist navigation in an intelligent navigation scenario as described above.

[0012] Alternatively, it includes a readable storage medium having a computer program stored thereon, and when the computer program is executed, it implements the method of using a knowledge graph to assist navigation in an intelligent navigation scenario as described above.

[0013] The present invention proposes a navigation framework that performs reasoning on the next optimal action for navigation based on the current target view, comgraph (common sense knowledge graph) knowledge and meta-learning. Specifically, in the navigation indoor environment, a unified common sense knowledge graph is constructed based on target detectors, network neural networks and crawlers; a semi-supervised inductive learning method GAT (Graph ATtention network) is used to effectively extract and update the graph; meta-learning is used to explore such a dynamic navigation graph that connects the visible and invisible environmental distributions; finally, the reinforcement learning policy value is decomposed, and the comgraph inference value and the policy value function are spliced ​​to provide better decision-making for reinforcement learning. The framework of the present invention shows improvements of approximately 28% and approximately 12% over the current SOTA (state-of-the-art) in success rate and SPL (success length weighted), respectively; the provided comgraph performs better than the Visual Genome knowledge graph in cross-scene generalization, and has a high performance in Hits@5 * and MRR * The figures for the first and second half of this year were respectively increased by about 56% and about 39%.

[0014] Compared with the prior art, the technical solution of the present invention has the following advantages:

[0015] 1) This method enables the agent to navigate in novel indoor environments without being affected by changes in scene layout. This navigation framework uses deep learning methods to construct a navigation scene knowledge graph (comgraph) that fuses multi-source data. This graph contains rich navigation semantics and successfully generalizes to new scenarios.

[0016] 2) Navigation knowledge embedding is composed of a Graph Attention Network (GAT) for knowledge representation and reasoning on the knowledge graph, and meta-learning for updating and applying knowledge in unknown scenarios. Using the GAT embedding to represent and extract knowledge graph features as part of the input state for reinforcement learning navigation effectively assists navigation. Meta-learning helps reinforce the loss gradient during training, thereby addressing the generalization problem in new scenarios.

[0017] 3) ResNet-18 is used to extract features from the observed image in the scene and to semantically represent the target Glove word vector. The image features, semantic features, and spatial information features are then combined into an MLP-LSTM neural network for memory storage and passed to reinforcement learning for state representation, enabling complete navigation.

[0018] 4) Based on the target position relationship in the navigation scene, it uses prior knowledge to infer the shortest path, thereby giving reinforcement learning optimization decisions. The inference value is used as part of the action sampling to guide reinforcement learning navigation and improve navigation generalization capabilities.

[0019] The solution of the present invention is simple and convenient to implement and has strong practicality. It solves the problems of low practicality and inconvenience in actual application existing in related technologies, can improve user experience, and has important market value. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the overall framework diagram of the knowledge graph construction and navigation usage according to the embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of constructing a navigation knowledge graph based on the faster-rcnn deep neural network in an embodiment of the present invention.

[0022] Figure 3 3 is a convergence curve diagram obtained by comparing the method of the embodiment of the present invention with other experiments. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0024] An embodiment of the present invention provides a method for using knowledge graphs to assist navigation in intelligent navigation scenarios, enabling intelligent agents to achieve excellent navigation results in new indoor navigation environments without being affected by changes in scene layout. The navigation framework first uses deep learning methods to construct a multi-source data fusion navigation scene knowledge graph (comgraph). The navigation knowledge embedding is composed of a graph attention network (GAT) for knowledge representation and knowledge reasoning on the knowledge graph and meta-learning for knowledge update and application in unknown scenarios. At the same time, a Resnet18 neural network is used to extract features from the observed images in the scene, and a target global word vector representation (Glove) is used for semantic representation. The image features, semantic features, and spatial information features are then spliced ​​into a multi-layer perceptron and a long short-term memory network (MLP-LSTM) for memory storage and passed to reinforcement learning for state representation, thereby performing complete navigation. This method uses prior knowledge to infer the shortest path based on the target position relationship in the navigation scene, thereby providing reinforcement learning with optimized decisions. The inference value is used as part of action sampling to guide reinforcement learning navigation and improve navigation generalization capabilities.

[0025] The key of the present invention is to provide a general common sense knowledge graph, as well as a knowledge representation and reasoning model combined with a graph attention mechanism. According to the multi-target relationship in the navigation scene, the prior knowledge of the general common sense knowledge graph is used to infer the target position, on the one hand as one of the representations of the spatial information state of reinforcement learning, and on the other hand as part of the reinforcement learning strategy value function. The construction and use of this knowledge graph can improve the navigation success rate of the intelligent agent in new scenes and enhance the overall navigation effect. The embodiment of the present invention provides a method for using knowledge graphs to assist navigation in an intelligent navigation scene, and proposes a general common sense knowledge graph, as well as a knowledge representation and reasoning model combined with a graph attention mechanism, as shown in the attached figure. Figure 1 :

[0026] The input is an image observed by the agent in the navigation scenario. Visual features of 1x2048 are extracted from the image using a residual network (e.g., ResNet18). Semantic features of the target are embedded into word vectors of 1x300. Spatial structural features of 1x92 are extracted using a graph attention network (GAT) based on the knowledge graph constructed by the present invention. A multi-layer perceptron is then used to concatenate the image, semantic, and spatial features, resulting in a concatenation size of 1x2440. A 512-bit LSTM (Long Short-Term Memory) network stores the training parameters for every three steps. A two-layer fully connected (FC) layer outputs the values ​​of the reinforcement learning (A3C) actor-critic (ACR) function, i.e., the estimated values ​​of the policy function and value function. Finally, meta-learning is used to store the gradient of the training loss function. Meta-learning uses a fully convolutional (FC) layer to store the training parameters every five steps for application to new training scenarios. This enhances generalization to new scenarios. State Knowledge uses the prior knowledge of the general common sense knowledge graph based on the multi-target relationship in the navigation scene to obtain the feature value of the optimal state of the next target through the multi-head attention mechanism and infer the target position. Its role is, on the one hand, to represent the spatial information state of reinforcement learning, and on the other hand, to serve as part of the value function of the reinforcement learning strategy. t ) represents the reinforcement learning policy function, v(o t ) is the value function, which is the total value function after the reinforcement learning value function estimation and the graph attention network value function estimation are spliced ​​together. R t is the reward value in the current state, O t+1It is the observed state obtained by the next action. The gradient flow represents the process of the neural network in the entire navigation framework through the back propagation direction of the loss function, gradient descent, and partial derivative. Forward propagation is the process of extracting features of different types of data through neural networks. The interactive flow is reinforcement learning, the process of interacting with the environment, and the input flow is the three main inputs of the model of the present invention, the environmental image features, the target word vector features and the features of the knowledge graph adjacency matrix. The domain knowledge of the scene is obtained by regular expression through knowledge distillation to obtain a 92*92 adjacency matrix, in which the vertices are composed of 92 of the most common scene targets, and the edges are 92*92 edges of the target relationships, such as left, right, on and other 7 relationships. These relationships represent the general rules of scene layout common sense, and the knowledge graph adjacency matrix is ​​embedded and represented by the GAT graph attention network. The construction and use of this knowledge graph can improve the navigation success rate for intelligent agents in new scenes and enhance the overall navigation effect.

[0027] The present invention uses deep learning methods to construct and use knowledge graphs and improve navigation performance. The specific method for constructing navigation knowledge graphs is:

[0028] The knowledge graph of the embodiment of the present invention is constructed based on multi-source information fusion in the AI2THOR environment, which consists of three main steps.

[0029] (1) Capture target relationships, such as color, size, shape, whether it is switched on or off, from the original AI2THOR dataset. This can be done using a crawler.

[0030] (2) Use the YoLoV4 object detector to generate relationships with its coordinates, generating up, down, left, right, inside, and contained relationships.

[0031] (3) Design a graph generation deep learning model as shown in the attached Figure 2 , predicting up, down, left, and right relations in scene graphs.

[0032] For ease of reference, the attached Figure 2 The specific implementation is as follows. You can set the corresponding parameters as needed during implementation:

[0033] The VGG16 neural network is preferably used to extract the spatial features of the depth map, i.e., the depth features, through 13 convolutional layers and 3 fully connected layers. The VGG16 neural network input image size is 224*224*3, and its 5 layers of max pooling layers, convolution layers, and activation functions all use ReLU functions. The final fully connected layer has 10,000 neurons.

[0034] Faster-RCNN is used to extract scene-target relationship features, and RPN (region proposal network) is used to generate corner boxes. The corner boxes are cropped and filtered, and then judged and corrected by the softmax map normalization layer to obtain more accurate recommended areas. A fixed feature map is generated through feature mapping and ROI (region of interest). The softmax function and L1 loss function are used for joint training to obtain accurate detection boxes x, y, w, h, i.e. center point coordinates, width and height. Figure 2 It is denoted as a two-dimensional detection box (x″, y″, w″, h″).

[0035] The agent's position and direction are encoded, concatenated with the output of faster-rcnn, and passed through a fully connected layer to obtain position features.

[0036] Finally, the depth features and position features are concatenated and passed through the fully connected layer to obtain the 3D detection frame x, y, z, w, h. Figure 2 It is recorded as a 3D detection box (x, y, z, weight, height).

[0037] The target to be found in navigation is matched with the detected target, the position relationship between the targets is calculated, and stored in the comgraph dataset.

[0038] In step (1), during the construction of the navigation knowledge graph of the embodiment of the present invention, multi-source fusion technology is used to capture most of the on and in relationships between objects and object containers, and about 40,000 texture triplets in real discrete environments are used as part of the background knowledge.

[0039] In step (2), the pre-trained YoLoV4 is applied and tested in the AI2THOR environment.

[0040] The present invention defines the relationship between left, right, over and under as follows:

[0041] left:(x j +w j / 2)-(x i +w i / 2)≥(w i +w j ) / 2& ①

[0042] 0≤|(y i +h i / 2)-(y j +h j / 2)|≤(h j +h i ) / 2 (②

[0043] right:(x i +w i / 2)-(x j +w j / 2)≥(w i +w j ) / 2& ③

[0044] 0≤|(y i +h i / 2)-(y j +h j / 2)|≤(h j +h i ) / 2 ④

[0045] over:(y i +h i / 2)-(y j +h j / 2)≥(h i +h j ) / 2& ⑤

[0046] 0≤|(x i +w i / 2)-(x j +w j / 2)|≤(w j +wi) / 2 ⑥

[0047] under:(y i +h i / 2)-(y j +h i / 2)≥(h i +h j ) / 2& ⑦

[0048] 0≤|(x i +w i / 2)-(x j +w j / 2)|≤(w j +w i ) / 2 ⑧

[0049] For each observation frame, the agent randomly walks in 120 rooms and detects several object relationships. The present invention stores 2700 triples. The object detector traverses every two objects i and j and predicts the relationship according to the above formula as follows:<Oi,relation,Oj> . Coordinate generation (xi, yi) = (boxes[i][0], boxes[i][1]); (wi, hi) = (boxes[i][0], boxes[i][1]). Where x i,y i , h i , w i Represents the 2D coordinates, height, and width of the center point of the i-th detection box, respectively. boxes represents the detection box. left, right, over, and under represent the relationship between the objects, left, right, up, and down, respectively.

[0050] In step (3), the embodiment of the present invention annotates 1.35 million images of four types of indoor rooms from AI2THOR, including 108 categories, and processes them using the relational common sense graph extraction module of the embodiment.

[0051] In this step, the specific implementation of the embodiment is described as follows:

[0052] (1) The embodiment uses faster-rcnn as a detector and pairs it with real objects to generate a comgraph. The connections between graph nodes are determined by their co-occurrence spatial relationship and the confidence of the object category provided by the object detector. It is worth noting that the embodiment does not use convolutional features from ROI pooling because there may be objects with various shapes and textures. In addition, the method of using ROI pooling makes the model difficult to generalize. Instead, the embodiment uses ROI pooling in space. 8 In the object O x Defines a feature f x , including center point coordinates (xcenter, ycenter), position (xmin, ymin, xmax, ymax), height, width and confidence value.

[0053] (2) Object O x and O j The intelligent agents are connected and the position features are obtained using the multi-layer perceptron (MLP). The deep image features are extracted using the VGG16 network.

[0054] (3) Example Matching Real Target O x Tags and Objects j By using the above three methods, the present invention can obtain 63,000 triples of relationships stored in comgraph. <s-p-o>Triples are used to construct the multi-source data fusion navigation scene knowledge graph comgraph of the present invention.

[0055] The knowledge graph G(S, P, O) of the present invention encodes the semantics and relevance of objects in the scene.

[0056] The general expression of a semantic relationship is <subject-prediction-object> between an object pair. It is worth noting that this semantic relationship is directional, that is, it connects one target (subject noun) to another target (object) through a relationship. The prediction can be an action or interaction between objects. Given an observed image I with subject i as the head entity and object j as the tail entity, the entity pair set (S i , O j ) represents the relationship between the i-th node of S head and the j-th node of T tail. k It is the relationship between entity pairs that represents the graph edges. All edges are initialized by weight settings.

[0057] Furthermore, the initial weights for edges in this embodiment of the present invention are set based on the TF-IDF ranking: (on: 0.163265, in: 0.204082, under: 0.122449, left: 0.102041, right: 0.081633, over: 0.142857, contain: 0.183673). Through these settings, the present invention considers relationships such as on and in to be more important, which can be trusted in navigation knowledge reasoning and accelerate the search for the target object.

[0058] For example, when the target apple is invisible, it is more likely to be in the refrigerator than to the left of the bowl or knife. This creates an accurate and universal commonsense knowledge graph for indoor navigation scenarios, laying the foundation for improved navigation results.

[0059] The specific methods used in the navigation knowledge graph are:

[0060] The knowledge graph overcomes the knowledge redundancy and inaccuracy that arise from different distributions across similar scenarios. In new scenarios, the present invention utilizes GAT to construct a knowledge graph update and adaptation framework. This is challenging because perfect prior knowledge applicable to all scenarios does not exist. Prior knowledge is considered an initial belief about a new scenario, rather than a strict rule. In a new scenario, the initial prior is insufficient due to the lack of goals or relationships. However, the spatial structure distribution of the initial knowledge graph for rooms of the same type is similar, and meta-learning can be used to change the implicit distribution. The specific principle is the second-order derivative information provided by one (or several) steps of the meta-learning model, which is called task curvature. This high-order information provides directional information for model initialization, namely the gradient direction for each task used by the present invention. The optimization is divided into two layers: an inner loop and an outer loop. The inner loop is the training process, which learns the basic skills for handling each task. The outer loop is the meta-training process that learns multi-task generalization capabilities. Using GAT, the present invention's intelligent agent can determine whether it is within an adjacent layer or across different scenarios by assigning different weights to neighboring nodes. In this way, the present invention can extract features from the comgraph according to different scenarios, thereby navigating more effectively. In addition, the comgraph structure is adopted in the meta-learning supervision stage. For the comgraph represented by the graph attention network GAT embedding of the present invention, the present invention has:

[0061]

[0062]

[0063] The present invention has learned A new node The output feature representation is obtained by aggregating the multi-head graph attention layers. During the training process, the edge and node operations are dynamically learned and then extracted and adjusted by the strategy. Using the aggregation function ||, the present invention connects K independent attention mechanisms by performing the transformation of ⑩. ij Represents the attention probability between node i and node j, w is the training parameter, W is the weight matrix, and σ() is nonlinear. N i k∈N in i Node is a neighbor of node i, x i is the attribute feature of node i. By performing edge operations on the adjacency matrix using the attention mechanism, GAT graph embedding enables the present invention to go beyond the integration of prior knowledge and move towards knowledge learning and graph adaptation. Only the main features of a partial graph need to be updated, without updating the features of the entire graph.

[0064] Experimental setup:

[0065] When experimenting with the method of the present invention, we used the A12THOR simulator as the experimental framework. It is divided into four different room types: bathroom, living room, bedroom, and kitchen. The experimental setting is SAVN. Each room is trained using 20 scenes.

[0066] 5 validation scenarios and 5 test scenarios. All methods are trained until the maximum convergence of 200 million frames.

[0067] Kitchen: microwave, toaster, refrigerator, coffee maker, trash can, bowls and boxes;

[0068] Living room: TV, pillows, trash can, laptop, bowls and boxes;

[0069] Bedroom: Indoor plants, lamps, books, and alarm clocks;

[0070] Bathroom: Toilet paper, soap bottle, light switch, sink, 21 in total.

[0071] The total number of available objects is 92, including the comgraph in the embodiment of the present invention. To train the model of the present invention, the embodiment of the present invention uses the Pytorch framework, uses RMSprop for adaptation optimization, and uses SharedRMSprop.

[0072] Experimental parameter settings:

[0073] To process visual perception images, this embodiment of the present invention uses a pre-trained ResNet18 to extract observation features at each time step. This method uses Glove to generate 300-dimensional semantic embeddings of the target and graph objects, for a total of 92 objects. The input of the reinforcement learning actor-critic network in this embodiment of the present invention is concatenated with the target object to form a 300-dimensional vector, the observation features are used as a 1024-dimensional feature vector, and the 92-node knowledge graph is input to the GAT to produce a 92-dimensional vector. The GAT is also used for knowledge reasoning that produces a single value.

[0074] The actor-critic network in this embodiment of the present invention consists of an LSTM network with 512 hidden states and two FC layers representing the actor-critic. The actor outputs a 6-dimensional action distribution, and the critic estimates a value using a softmax normalized mapping activation function.

[0075] The value after GAT decomposition is input into the criticism embedding to form a value estimate. In particular, in terms of dynamically updating the knowledge graph and correcting the error prior of the policy network in unknown scenarios, the embodiment of the present invention adopts a new meta-learning MAML agent. The input of the graph is a 1024-dimensional vector as a node feature. This method uses the glove embedding method to map 512 observation features to 300 to 512 targets through a linear layer. Each layer contains 92 adjacency matrix nodes, for a total of 5 layers, of which 4 layers are edges between objects in four scenarios, and the other layer is regularized using a self-connection layer. The training results of the entire model are shown in the attached figure. Figure 3 As shown, compared with the results obtained by other existing methods such as A3C, the use of GAT-SAVN, that is, the model of the embodiment of the present invention, has the best navigation efficiency, indicating that the framework of the present invention can find the target faster. At the same time, the more the knowledge graph is used, the higher the model navigation efficiency.

[0076] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0077] In some possible embodiments, a system for using knowledge graphs to assist navigation in an intelligent navigation scenario is provided, including a processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute a method for using knowledge graphs to assist navigation in an intelligent navigation scenario as described above.

[0078] In some possible embodiments, a system for using a knowledge graph to assist navigation in an intelligent navigation scenario is provided, comprising a readable storage medium having a computer program stored thereon. When the computer program is executed, a method for using a knowledge graph to assist navigation in an intelligent navigation scenario as described above is implemented.

[0079] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A method for using knowledge graphs to assist navigation in intelligent navigation scenarios, characterized by: First, we use deep learning methods to build a multi-source data fusion navigation scenario knowledge graph comgraph, which is used to extract features from the comgraph according to different scenarios to support navigation. The Graph Attention Network (GAT), which represents and infers knowledge on the knowledge graph, and meta-learning, which updates and applies knowledge in unknown scenarios, form navigation knowledge embedding to extract spatial structural features. A neural network is used to extract image features from the observed images in the scene, and semantic features are extracted through the target global word vector representation. The image features, semantic features, and spatial structural features are then spliced ​​into a multi-layer perceptron and a long short-term memory (MLP-LSTM) network for memory storage, and then passed to reinforcement learning for state representation, enabling complete navigation. The graph of the graph attention network GAT embedding representation is as follows: Among them, learning After that, a new node The output feature representation is obtained by aggregating multiple graph attention layers; During training, edge and node operations are learned dynamically and then extracted and adjusted by the policy; using the aggregation function ||, K independent attention mechanisms are connected by performing the transformation of formula (2); v ij represents the attention probability between node i and node j, w is the training parameter, W is the weight matrix, σ() is nonlinear; N i k∈N in i Node is a neighbor of node i, x i It is the attribute characteristic of the i-node.

2. The method for using knowledge graphs to assist navigation in an intelligent navigation scenario according to claim 1, characterized in that: Based on the use of deep learning methods to construct a multi-source data fusion navigation scene knowledge graph, the navigation semantic knowledge is generalized to new scenes, so that the intelligent agent is not affected by changes in scene layout in the new indoor navigation environment.

3. The method for using knowledge graphs to assist navigation in an intelligent navigation scenario according to claim 2, characterized in that: When the graph attention network GAT and meta-learning form navigation knowledge embedding, the GAT embedding is used to represent and extract knowledge graph features as the input state for reinforcement learning navigation; Meta-learning helps reinforce the loss gradient of learning during training and achieve generalization in new scenarios.

4. The method for using knowledge graphs to assist navigation in an intelligent navigation scenario according to claim 1, characterized in that: The Resnet18 neural network is used to extract image features from the observed images in the scene.

5. The method for using knowledge graphs to assist navigation in an intelligent navigation scenario according to claim 1, characterized in that: The image features, semantic features, and spatial structure features are spliced ​​into the MLP-LSTM neural network for memory storage and passed to reinforcement learning for state representation for complete navigation.

6. The method for using knowledge graphs to assist navigation in an intelligent navigation scenario according to claim 1, characterized in that: Based on the target position relationship in the navigation scene, the shortest path is inferred using prior knowledge, thereby giving reinforcement learning optimization decisions. The inference value obtained from the inference of the shortest path is used as part of the action sampling to guide reinforcement learning navigation and improve navigation generalization ability.

7. A system for using knowledge graphs to assist navigation in intelligent navigation scenarios, characterized by: A method for implementing knowledge graph-assisted navigation in an intelligent navigation scenario as described in any one of claims 1-6.

8. A system for assisting navigation using a knowledge graph in an intelligent navigation scenario, characterized by: It includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a method for using knowledge graphs to assist navigation in an intelligent navigation scenario as described in any one of claims 1-6.

9. A system for assisting navigation using a knowledge graph in an intelligent navigation scenario, characterized by: It includes a readable storage medium, on which a computer program is stored. When the computer program is executed, it implements a method for using a knowledge graph to assist navigation in an intelligent navigation scenario as described in any one of claims 1 to 6.

Images