CN118193628A - Semantic enhancement graph contrast learning method and system for road network representation - Google Patents

Abstract

The invention relates to a semantic enhancement map contrast learning method and system for road network representation, and belongs to the technical field of traffic data analysis and machine learning. The invention provides a semantic enhancement map contrast learning framework, which effectively combines road attributes and visual information through multi-mode feature embedding, and deeply mines and utilizes the features of road network data. A novel road network data representation is constructed, so that the road network data representation not only has rich information with high dimensionality, but also can be directly used as input of a machine learning model. The design makes various downstream tasks based on road network, such as road label classification, speed prediction and the like, more convenient to implement. The scheme integrates the ideas of graph comparison and learning, improves the robustness and generalization capability of road network data representation by simulating the topology change and data loss of the road network, effectively solves the limitation of the traditional method in complex road network analysis, and improves the road network data processing efficiency.

Description

Semantic enhancement graph contrast learning method and system for road network representation

Technical Field

The invention belongs to the technical field of traffic data analysis and machine learning, and relates to a semantic enhancement map contrast learning method and system for road network representation.

Background

With the acceleration of urban traffic networks becoming increasingly complex and diverse, traditional road network analysis methods face increasing challenges. Traditional road network analysis methods often rely on simplified mathematical models and limited data sources, making it difficult to capture and interpret complex dynamics and diversity in the road network. For example, in urban planning and traffic management, deep understanding of key information such as traffic flow, congestion patterns, accident risks, etc. is of paramount importance, but the existing technologies and methods often cannot effectively process and analyze large-scale and high-dimensional road network data. In addition, with the development of technology, the types and magnitudes of road network data have changed significantly. The richness and diversity of data is unprecedented from the GPS track of the vehicle to the monitoring of real-time traffic conditions. However, this also presents new challenges for data processing and analysis. Existing analysis tools and methods often take care of handling such large data, especially in situations where it is desirable to monitor traffic flow on-the-fly, predict traffic trends, and address traffic emergency situations. For example, in terms of real-time traffic condition monitoring, conventional data analysis methods cannot reflect changes in traffic flow instantaneously; in the aspect of road condition prediction, due to lack of deep analysis on big data, future traffic conditions cannot be predicted accurately; in the aspect of traffic accident response, the traditional method also has obvious defects in the aspects of timeliness and accuracy of data processing. Thus, for the complex and diverse nature of modern urban traffic networks, there is an urgent need to develop new analytical methods and techniques to better understand and manage these complex traffic systems. These new methods and techniques should be able to effectively cope with large-scale, high-dimensional traffic data to achieve in-depth understanding and accurate management of traffic networks. This includes not only efficient processing of existing data, but also efficient modeling and prediction of traffic network topology changes and data loss conditions. Through the technical progress, urban planning and traffic management can be supported more comprehensively and accurately, so that the efficiency and the safety of an urban traffic system are improved.

Disclosure of Invention

In view of the above, the present invention aims to provide a semantic enhancement map contrast learning method and system for road network representation, which provides a framework named SE-GCL (semantic enhancement map contrast learning), and the framework effectively combines road attribute and visual information through multi-modal feature embedding, and deeply digs and fully utilizes the features of road network data. The novel road network data representation is constructed, not only has rich information with high dimensionality, but also can be directly used as input of a machine learning model. The design makes various downstream tasks based on road network, such as road label classification, speed prediction, traffic time prediction and the like, more convenient to implement. SE-GCL integrates the ideas of graph comparison and learning, and improves the robustness and generalization capability of road network data representation by simulating the topology change and data loss of a road network, thereby effectively solving the limitation of the traditional method in complex road network analysis.

In order to achieve the above purpose, the present invention provides the following technical solutions:

the method for the semantic enhancement map contrast learning of the road network representation constructs a semantic enhancement map contrast learning framework for the road network representation, and specifically comprises the following stages:

the first stage: constructing a multi-mode feature embedding module, which respectively obtains attribute embedding (embedding, namely low-dimensional representing vector) and visual feature embedding of the road section by utilizing the attribute data and street view image data of the road section, and splicing the attribute embedding and the visual feature embedding together to serve as initial embedding of the road section;

and a second stage: constructing a graph augmentation module, designing two different augmentation strategies on the basis of modeling a road network into a graph, and respectively applying disturbance to topology and node characteristics of the road network (graph), thereby generating two different augmentation graphs (graph varieties);

And a third stage: constructing a graph representation learning module, and extracting node levels, namely the representation of road sections, by adopting a graph neural network model on the basis of two graph varieties generated by the graph augmentation module;

fourth stage: a contrast optimization module based on semantic enhancement is constructed, sampling of sample pairs is carried out by utilizing road network characteristics and mobile semantics in a historical track, and a contrast loss function of double views is designed to guide a contrast optimization process.

Further, in the first stage, the method is mainly divided into an attribute feature embedding part and a visual feature embedding part;

The attribute feature embedding section includes: for any road section v _i, the attributes of the road section identifier ID, the road section length LEN, the road section midpoint coordinates (LON, LAT) and the like are synthesized and analyzed, and the signs corresponding to the attributes are v _i.id、v_i.len、v_i.lon、v_i and LAT respectively; considering that the change of longitude and latitude coordinates in a single road section is very limited, only the longitude and latitude of the geometric midpoint of the road section is selected to represent the position information of the road section. For real type attributes v _i. Lon and v _i. Lat, converting them into a discretized form through a binning operation, then mapping each attribute value into a corresponding feature vector using a separate linear layer, and then stitching together the four feature vectors to form the attribute embeddings of the road segments;

the visual characteristic embedding section includes: the pre-trained Swin transducer is used as a visual characteristic encoder, panoramic image data collected at the midpoint of a road section is used as input, and visual embedding of the panoramic image data is output Finally, two different feature types are connected together to obtain the initial embedding of the road segment.

Further, in the second stage, a graph augmentation module is constructed to augment the road network by using two strategies of mobility-based edge removal and modality-based feature coverage; the mobility-based edge removal strategy aims at the topological structure of the road network, and the edge removal probability is determined according to the transition probability among road sections in the historical track data; if two road segments occur together often in a historical track, the connection between them is considered strong and important, so the probability of removing such a connection will be low; the feature masking strategy based on the mode is a disturbance strategy aiming at node features, and one of the attributes or the image features is randomly selected to mask so that the contrast learning model is adapted to the loss of road section data, and therefore the robustness of the model is enhanced; during each training process, two image views are generated simultaneously for comparison by using the two different augmentation strategies.

Further, in the third stage, a graph representation learning module is constructed, and representations of the two graph views are respectively learned by taking the two graph views generated by the graph augmentation module as inputs; the method specifically comprises the following steps: advanced features are learned and extracted from the graph view using a graph encoder to generate a complex representation of each segment, and finally, the segment representations generated by the graph encoder are mapped to a low-dimensional potential space by a nonlinear projection head.

Further, in the fourth stage, a contrast optimization module based on semantic enhancement is constructed, wherein the module comprises a semantic enhancement sampling strategy, and positive and negative samples can be constructed by utilizing road network characteristics and mobile semantics in a historical track; the method specifically comprises the following steps: given a target node (i.e., an anchor node) of a certain view, a neighbor node in the h-hop is regarded as an intra-view positive sample, and a mirror node of another view is regarded as an inter-view positive sample; for nodes beyond the h-hop range, further screening out nodes co-occurring with anchor nodes in the history track, and selecting the rest nodes as real negative samples; a contrast loss function based on InfoNCE target design was then used:

to calculate the contrast loss of all nodes on the two graph views and obtain the final dual-view contrast loss function

Wherein v' _i represents the mirror node of v _i in another view; finally, training of the network is performed using the loss function to optimize the road network embedded representation.

The invention also provides a semantic enhancement map contrast learning system for road network representation.

The invention has the beneficial effects that:

The invention provides a framework named SE-GCL (semantic enhanced graph contrast learning), which effectively combines road attributes and visual information through multi-mode feature embedding, and deeply mines and fully utilizes the features of road network data. The novel road network data representation is constructed, not only has rich information with high dimensionality, but also can be directly used as input of a machine learning model. The design makes various downstream tasks based on road network, such as road label classification, speed prediction, traffic time prediction and the like, more convenient to implement. SE-GCL integrates the ideas of graph comparison and learning, and by simulating the topology change and data loss of the road network, the robustness and generalization capability of road network data representation are improved, the limitation of the traditional method in complex road network analysis is effectively solved, and the efficiency of road network data processing and the understanding of complex dynamics of the urban traffic network are improved.

Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objects and other advantages of the invention may be realized and obtained by means of the instrumentalities and combinations particularly pointed out in the specification.

Drawings

For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in the following preferred detail with reference to the accompanying drawings, in which:

FIG. 1 is a system block diagram of the present invention;

FIG. 2 is an illustrative example diagram of a semantic enhanced sampling strategy.

Detailed Description

The technical scheme of the invention is described in detail below with reference to the accompanying drawings.

The semantic enhancement map contrast learning framework for road network representation provided by the invention is used for representation learning of the road network. Currently, the increasing complexity of urban traffic networks results in massive and diverse road network data, the processing and analysis of which entails huge storage and computational overhead. Currently, road network data is typically processed by combining road attributes and visual information. The invention provides a new framework for the representation and optimization of the data, which comprises the following four stages: firstly, constructing a multi-mode feature embedding module, and obtaining attribute and visual feature embedding of a road section by utilizing attribute data and street view image data of the road section; secondly, two kinds of augmentation strategies are designed to apply disturbance to the topology and node characteristics of the road network through a graph augmentation module, and two different augmentation graphs are generated; next, in a graph representation learning module, extracting representations of road segments from the two graph variants using a graph neural network model; finally, in a contrast optimization module based on semantic enhancement, sample pair sampling is carried out by combining road network characteristics and mobile semantics in a history track, and a double-view contrast loss function is designed to guide the optimization process. The framework improves the efficiency of road network data processing and understanding of the complex dynamics of the urban traffic network. Fig. 1 is a system block diagram of the present invention.

In the present embodiment, 7 necessary concepts are referred to as follows:

Concept 1: road network, which is a directed graph g= (V, a). Wherein each vertex V e V represents a segment of a road that contains attributes such as length and coordinates on the map. A is an adjacency matrix in which each entry a _i,j is a binary value indicating whether road segment v _j is connected to road segment v _i.

Concept 2: street view imageIs a photograph taken of a vehicle at a particular location on a road network. Specifically, we collect multiple images for each road segment and stitch them into a 360 degree panoramic view. Panoramic images represent a snapshot illustrating the visual details of the surrounding environment.

Concept 3: trackIs a sequence of GPS points recorded as the vehicle moves over the road network, noted < p ₁,p₂,p₃,p₄,...,p_|T| >. Where p _i is the coordinates of the ith point and |T| represents the total number of points. We further map each track onto the road network using the L2MM depth model we propose. Thus, the trajectory is represented by a series of connected road segments < v ₁,v₂,v₃,v₄,...,v_m > in the road network G.

Concept 4, z _i refers to the embedding or representation of a particular node (e.g., road segment). In a graph, each node has a unique embedding to characterize it.Representing the embedding of positive samples. In contrast learning, a positive sample is a node that is similar to or associated with the anchor node (z _i). /(I)Representing the embedding of the negative sample. In contrast to the positive samples, the negative samples are nodes that are dissimilar or unrelated to the anchor node.

The 5 th concept, the anchor node, refers to a reference point or target node for comparison. During the learning process, the representation (or feature vector) of the anchor node is typically compared with representations of other nodes (positive or negative samples) to learn the discriminative node embedding.

Concept 6, h-hop, is used to describe the neighborhood of nodes in the graph. In a graph, an h-hop neighborhood of a node refers to the set of all nodes that can be reached by not more than h hops from that node.

The 7 th concept, infoNCE target a loss function that is commonly used for contrast learning. The basic idea is to maximize the similarity between positive pairs of samples while minimizing the similarity between negative pairs of samples. In contrast learning, positive sample pairs are typically similar or related entities, while negative sample pairs are uncorrelated entities.

FIG. 2 is an illustrative example diagram of a semantic enhanced sampling strategy, the system framework of the present system mainly includes four modules: the system comprises a multi-mode feature embedding module, a graph augmentation module, a graph representation learning module and a semantic enhancement-based contrast optimization module, wherein a system block diagram of the invention is shown in fig. 1, and the system block diagram comprises: a multi-modal feature embedding module: the module starts from obtaining attribute embedding and visual feature embedding of the road section by utilizing the attribute data and street view image data of the road section, and then combining the attribute embedding and the visual feature embedding to serve as initial embedding of the road section, and comprises the following steps:

step one: input road network g= (V, a), street view image data Historical GPS track data/>Obtaining attribute embedding X ^a and visual feature embedding X ^v of the road section through a multi-mode feature embedding module;

Step two: and splicing the attribute embedding and the visual feature embedding of the road segments to construct an initial embedding X=X ^a⊕X^v of the road segments, and taking the initial embedding X=X ^a⊕X^v as a node (road segment) feature matrix of the road network.

And a graph augmentation module: the module applies disturbances to the topology and node characteristics of the road network (graph) respectively using two different augmentation strategies, thereby generating two different augmentation graphs (graph variants), comprising the steps of:

step one: calculating edge deletion probability according to road historical track data Where e _i,j denotes an edge connection between road segments v _i and v _j,/>Representing a transition probability estimate from road segment v _i to v _j;

step two: according to the edge deletion probability Pr and the road network G as input, disturbance is applied to the road network twice (each disturbance relates to two aspects of topology and node characteristics), and a first augmentation chart is obtained Its adjacency matrix is/>The feature matrix is/>Second augmentation chart/>Its adjacency matrix is/>The feature matrix is/>

The diagram represents a learning module: the starting point of the module is that on the basis of two graph varieties generated by the graph augmentation module, a graph neural network model is adopted to extract node levels, namely the representation of road segments, and the method comprises the following steps:

Step one: adjacency matrix for a first augmented graph obtained in a graph augmentation module using a graph encoder Feature matrix/>Performing coding operation to obtain an embedded matrix/>

Step two: adjacency matrix for a second augmented graph obtained in a graph augmentation module using a graph encoderFeature matrix/>Performing coding operation to obtain an embedded matrix/>

Step three: and mapping the two embedded matrixes obtained by the graph encoder to a low-dimensional potential space through a nonlinear projection head to obtain embedded Z ₁ and Z ₂ of two groups of road sections.

A contrast optimization module based on semantic enhancement: the module starts from sampling a sample pair by utilizing road network characteristics and mobile semantics in a history track, designing a double-view contrast loss function to guide a contrast optimization process and training a network by using the loss function so as to optimize an embedded representation of a road network, and comprises the following steps:

step one: and constructing a co-occurrence matrix M according to the road history track T. Then, based on the co-occurrence matrix M and the number h of neighboring nodes, adjacent matrixes of the two augmentation graphs are respectively obtained And/>Sampling of positive and negative samples is performed. Specifically, given a target node (i.e., anchor node) of a view, neighbor nodes within its h-hops are considered intra-view positive samples, and mirror nodes of another view are considered inter-view positive samples. For nodes except the h hops, further excluding nodes co-occurring with the anchor node in the history track, and selecting the rest nodes as real negative samples;

step two: using a InfoNCE-target-design-based dual-view contrast loss function as an input according to the sampling result of the step one And calculating a loss value and training the network by taking the loss value as a guide.

Finally, it is noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention may be modified without departing from the spirit and scope of the technical solution, and all such modifications are included in the scope of the claims of the present invention.

Claims (6)

1. A semantic enhancement map contrast learning method for road network representation is characterized in that: the method constructs a semantic enhancement graph contrast learning framework for road network representation, and specifically comprises the following stages:

The first stage: constructing a multi-mode feature embedding module, respectively obtaining attribute embedding and visual feature embedding of the road section by utilizing the attribute data and street view image data of the road section, and splicing the attribute embedding and the visual feature embedding of the road section to serve as initial embedding of the road section;

and a second stage: constructing a graph augmentation module, designing two different augmentation strategies on the basis of modeling a road network as a graph, and respectively applying disturbance to topology and node characteristics of the road network so as to generate two different augmentation graphs;

And a third stage: constructing a graph representation learning module, and extracting node levels, namely the representation of road sections, by adopting a graph neural network model on the basis of two graph varieties generated by the graph augmentation module;

fourth stage: a contrast optimization module based on semantic enhancement is constructed, sampling of sample pairs is carried out by utilizing road network characteristics and mobile semantics in a historical track, and a contrast loss function of double views is designed to guide a contrast optimization process.

2. A semantic enhanced graph contrast learning method for road network representation according to claim 1, characterized by: in the first stage, the method mainly comprises attribute feature embedding and visual feature embedding;

The attribute feature embedding section includes: for any road section v _i, the road section identifier ID, the road section length LEN and the road section midpoint coordinate (LON, LAT) attributes are integrated for analysis, and the signs corresponding to the attributes are v _i.id、v_i.len、v_i.lon、v_i and LAT respectively; for real type attributes v _i. Lon and v _i. Lat, converting them into a discretized form through a binning operation, then mapping each attribute value into a corresponding feature vector using a separate linear layer, and then stitching together the four feature vectors to form the attribute embeddings of the road segments;

the visual characteristic embedding section includes: the pre-trained Swin transducer is used as a visual characteristic encoder, panoramic image data collected at the midpoint of a road section is used as input, and visual embedding of the panoramic image data is output Finally, two different feature types are connected together to obtain the initial embedding of the road segment.

3. A semantic enhanced graph contrast learning method for road network representation according to claim 2, characterized by: in the second stage, a graph augmentation module is constructed to augment the road network by using two strategies of mobility-based edge removal and modality-based feature coverage; the mobility-based edge removal strategy aims at the topological structure of the road network, and the edge removal probability is determined according to the transition probability among road sections in the historical track data; if two road segments occur together often in a historical track, the connection between them is considered strong and important, so the probability of removing such a connection will be low; the feature masking strategy based on the mode is a disturbance strategy aiming at node features, and one of the attributes or the image features is randomly selected to mask so that the contrast learning model is adapted to the loss of road section data, and therefore the robustness of the model is enhanced; during each training process, two image views are generated simultaneously for comparison by using the two different augmentation strategies.

4. A semantic enhanced graph contrast learning method for road network representation according to claim 3, characterized by: in the third stage, a graph representation learning module is constructed, and the two graph views generated by the graph augmentation module are taken as inputs to respectively learn the representation of the two graph views; the method specifically comprises the following steps: advanced features are learned and extracted from the graph view using a graph encoder to generate a complex representation of each segment, and finally, the segment representations generated by the graph encoder are mapped to a low-dimensional potential space by a nonlinear projection head.

5. The semantic enhanced graph contrast learning method for road network representation according to claim 4, wherein: in the fourth stage, a contrast optimization module based on semantic enhancement is constructed, wherein the module comprises a semantic enhancement sampling strategy, and positive and negative samples can be constructed by utilizing road network characteristics and mobile semantics in a historical track; the method specifically comprises the following steps: giving a target node of a certain view, regarding a neighbor node in the h-hop as an intra-view positive sample, and regarding a mirror node of another view as an inter-view positive sample; for nodes beyond the h-hop range, further screening out nodes co-occurring with anchor nodes in the history track, and selecting the rest nodes as real negative samples; a contrast loss function based on InfoNCE target design was then used:

to calculate the contrast loss of all nodes on the two graph views and obtain the final dual-view contrast loss function

Wherein v' _i represents the mirror node of v _i in another view; finally, training of the network is performed using the loss function to optimize the road network embedded representation.

6. A semantic enhanced graph contrast learning system for road network representation, characterized by: the system employs the method of any one of claims 1 to 5.