Method and an apparatus for correcting junction shapes in a road network map

Abstract

For obtaining more realistic junction shapes in a road network map, a classifier (212) obtains crosswalk features from an aerial image (208, 308) of a road and an authoritative road network map (210, 310) of the road. An enhanced road network map (320) is created by co-locating the crosswalk features as crosswalk instances (312) into the authoritative road network map. Each crosswalk instance in the enhanced road network map includes coordinates of a boundary (316) of a single crosswalk. The crosswalk instances of the enhanced road network map are grouped into one or more junction instances by a geometrical reasoner, and their coordinates are refined based on coordinate centers of the junctions. Road network nodes, road network segments and attributes of the enhanced road network map are then updated to provide its connectivity. Afterwards, the enhanced road network map can be converted into a pre-defined format for road traffic simulations.

Description

[0001] METHOD AND AN APPARATUS FOR CORRECTING JUNCTION SHAPES IN A ROAD NETWORK MAP

[0002] TECHNICAL FIELD

[0003] The disclosure relates generally to a road network junction correction system from aerial images, and more particularly, the disclosure relates to a method for correcting junction shapes in a road network map. Moreover, the disclosure relates to an apparatus for correcting junction shapes in the road network map.

[0004] BACKGROUND

[0005] Mapping road networks remains a challenging and time-consuming task, heavily reliant on manual labor. Many countries face inadequate road map coverage outside urban centers, while the existing methods for constructing and updating street map datasets are impractical due to the high error rates of existing algorithms. One particularly significant challenge in road network generalization is accurately identifying and reproducing complex junctions in the input road network for simulators. The existing methods, which rely on statistical analysis and topological properties to characterize junction topology, are prone to errors. They fail to account for the structural complexity of the junctions, are sensitive to scale, and lack generality. In reality, complex junctions in the real-world exhibit significant variations in node density, enclosed shape and size, and structural form.

[0006] Additionally, an authoritative road network map typically utilized for traffic simulation comprises a simplified network composed of polygons or points of interest. However, this authoritative road network map requires intricate and standardized geometrical transformations to be integrated into a simulator. The authoritative road network map is commonly known as a "shapefile," which is a digital vector format designed for storing geographic location and associated attribute information. The primary issue with this format is its inability to store topological information. In the shapefile, unidirectional roads are represented as lines (i.e., road network edges), while intersections are represented as points (i.e., road network nodes) on the latitude / longitude plane. Additional metadata, such as speed limits, number of lanes, width and so on, may be included as supplementary attributes, if available. In this configuration, the road network nodes, typically positioned at the intersection points of different road network edges, are utilized to identify the primary junctions within the road network. The shape and extension of these junctions, including the intersection coordinates and the trajectory of internal lanes within the intersection, are represented as one or more one-dimensional points.

[0007] Furthermore, the authoritative road network maps are typically encoded using geo-coordinates, which necessitate conversion to the Cartesian coordinate system employed by the simulator through various projections. Occasionally, the coordinate reference system conversion between the satellite image and the reference map format can pose challenges. For instance, in China, a special obfuscating algorithm is employed to randomly offset latitude and longitude information, resulting in an inaccurate alignment between street maps and satellite imagery.

[0008] Various computer vision algorithms and approaches have been employed to retrieve not only road pavement but also lane markings, zebra crossings, buildings, and terrain characteristics. However, even the most advanced methods and models, capable of utilizing both Lidar and camera imagery to generate structures, still exhibit errors ranging from 5% to 10%. These significant error rates pose a considerable challenge for navigating road maps, especially as the number of roads increases. The high frequency of errors makes it virtually impossible to rely on such maps, leading to a decline in performance. Moreover, the definition of post-processing steps to address these errors can be error-prone in itself. Additionally, the execution time of search algorithms within the road graph tends to be slower, further impeding the efficiency of the mapping process.

[0009] Automated conflation systems and techniques have emerged as a solution for integrating vector and map imagery, specifically for road vector data, street maps, and ortho imagery. These systems aim to combine the realistic and accurate visuals of imagery with the detailed attribution information found in the original map representation. The primary driving factors behind the conflation process are the increasing demand for compiling a large number of digital maps within shorter timeframes and the rapid advancements in interactive management processes and mathematical algorithms designed to efficiently handle geographical data. Nevertheless, integrating geographical information from various sources and scales, while ensuring precision and consistency, presents a set of challenges. These challenges include transferring attributes, adding missing features, and reconciling heterogeneous data originating from different sources but covering the same geographical area. Consequently, addressing these challenges requires tailored solutions that are specific to the analyzed dataset, aiming to extract the most appropriate geometry and semantic data to create a newly formed geographic dataset. A Global Positioning System, GPS, trajectory refers to a sequence of GPS positions recorded as a vehicle traverses the road network during a single trip. By applying kernel density estimation, clustering, and trajectory merging algorithms to datasets comprising trajectories from multiple vehicles and trips, it becomes possible to generate a road map that automatically extends existing maps. However, the main challenges associated with GPS trajectories are twofold. Firstly, selecting an appropriate constant density threshold is crucial to optimize the balance between spurious and missing roads. This decision influences the accuracy of the resulting road map. Secondly, there is a lack of a standardized method for quantitatively evaluating graphs and maps in terms of their geographical location.

[0010] Known is a system for detecting roadway lane lines based on probability values assigned to a subset of pixels in an aerial image. In this system, a classifier (e.g., a deep learning model) generates a probability map that assigns probability values to at least a portion of the pixels. By comparing these values, the geographical location of roadway lane lines is determined. This approach addresses issues related to traffic occlusions from the vehicle's perspective. However, it encounters errors stemming from the co-location of lane lines and the hypothesis-linking algorithm used in post-processing. Another challenge lies in the precision of lane line detection and the accuracy of the training data labels (e.g., pixel coordinates for locations of lane markings).

[0011] Another existing solution uses an end-to-end cascaded convolutional neural network, CNN, architecture comprising two jointly trained components. The first component is responsible for detecting pixels on the road, while the second neural network focuses on pixels close to the road centerline. Subsequently, they apply a thresholding and thinning process to the centerline segmentation output, enabling the extraction of a graph. Their implementation demonstrates the ability to handle complex backgrounds and occlusions caused by trees and cars. Similarly, yet another existing solution proposes an approach to enhance the segmentation output using a conditional generative adversarial network. The segmentation CNN is trained not only to produce ground truth labels (using a mean-squared-error loss) but also to deceive a discriminator CNN that is trained to differentiate between the ground truth labels and the segmentation CNN outputs. This methodology has shown robustness against occlusions caused by cars and trees. However, it may not perform well in scenarios involving continuous large areas of occlusion. Yet another existing solution, called DeepRoadMapper methodology, introduces an additional post-processing step to infer missing connections within the initial extracted road network. Candidate missing connections are generated by conducting the shortest path search on a graph defined by the segmentation probabilities. Subsequently, a separate CNN is trained to accurately identify the correct missing connections. In practice, complex post-processing procedures are often necessary to interpret the segmentation outputs and extract topological structures required for map construction. Contrasting the approach of segmenting imagery, other existing solutions, called RoadTracer and Iterative Deep Learning (IDL) for Network Topology Extraction, utilize an iterative graph construction, IGC, approach, which extracts roads through a step-by-step search process. At each step, a CNN is queried to determine a direction in which to proceed with the search and a road segment is added to a partial road network graph in that direction. Although IGC methods enhance accuracy, they are significantly slower in execution time compared to segmentation approaches, rendering them unsuitable for interactive settings. Additionally, defining the post-processing steps in IGC- based methods is prone to errors. Overall, the primary challenge associated with all the known road mapping solutions is that while CNN architectures can achieve high performance in terms of the loss function on training examples, their performance during inference, or real-world application, tends to be poor.

[0012] Various solutions have been developed for an automated extraction of buildings from extensive satellite images, including both semantic and instance segmentation approaches. For instance, known is a solution that proposes a Multi-task U-Net network model capable of learning multiple detectors, each dedicated to a specific building size. This model is trained using knowledge distillation, leveraging another road extraction model as a teacher. In postprocessing, the outputs for each building size are integrated after binarization and filtering, following the common practice in semantic segmentation. Another existing solution employs an instance segmentation method that combines Mask R-CNN with building boundary regularization. This method directly produces polygon points for individual buildings, exhibiting better-regularized polygon boundaries compared to classical Mask R-CNN algorithms in terms of accuracy and completeness. However, obtaining regularized polygons that are directly applicable to cartographic and engineering problems remains a challenge in segmentation applications. In the case of semantic segmentation, the instantiation process involves a complex post-processing pipeline and requires specific knowledge of the analyzed network. Another existing solution provides for zebra crossing classification and detection, which plays a crucial role in various applications, such as mobility autonomy and road quality assessment. This task presents several challenges, including the creation of automatically annotated data considering occlusions, shadows, and other factors, limited generalization capabilities and recall of the final algorithm, and specific limitations for camera-based systems, such as maximum distance and minimum time constraints for crosswalk retrieval.

[0013] Yet another existing solution introduces a system capable of automatically acquiring and annotating satellite imagery of zebra crossings, training deep learning models for large-scale crosswalk classification. The system enables automated crosswalk annotation globally, benefiting visually impaired systems, autonomous driving technologies, and other related applications. Yet another existing solution presents a model for crosswalk detection and localization using satellite images from Google Maps to assist visually impaired individuals. The approach utilizes a Support Vector Machine, SVM, classifier combined with Google Road Map data to enhance computational efficiency and reduce false alarms. Yet another existing solution proposes an SVM algorithm for zebra crossing detection in aerial imagery. However, their datasets are relatively small, resulting in limited generalization capabilities. Furthermore, the use of phone cameras or car cameras has been explored to aid visually impaired individuals and driver assistance systems. Nevertheless, both solutions face limitations, particularly in terms of the maximum distance within which crosswalks can be reliably detected.

[0014] The identification of junctions within a road network has proven to be a challenging task, as the existing solutions have highlighted the difficulties in precisely characterizing and delineating these junctions. Most existing solutions have specifically focused on the retrieval of complex junctions, which exhibit structural complexity, where roads meet and join in intricate ways, and might have irregular geometric boundaries. Such methods are prone to errors, scale-dependency, and lack of generality within standardized road network mapping. For instance, complex junctions can be defined as locations with dense nodes of road intersections, and an existing solution proposes a density analysis approach to detect them. In that way, by carefully adjusting the density threshold, complex junctions, represented as clusters of densely connected nodes, could be identified using a combination of spatial clustering and graph theory. However, this approach may inadvertently include into consideration street blocks that are not part of the actual complex junction. In such cases, a spatial dendrogram derived from a given road junction could provide flexibility in creating alternative credible connectivity solutions for the network by selectively removing edges and associated vertices (edge restriction) or replacing vertices with a single vertex representing combined adjacencies (vertex contraction).

[0015] Yet another existing solution observes that node density alone did not consistently work in all cases. It utilizes statistics of redundant circles (also referred to as loops in topological terms) to characterize different types of complex junctions and attempted to detect them by searching for such circles within a road network. A similar approach is applied to the detection of highway interchanges, with a focus on identifying forks and y-nodes that are characteristic of such structures. Similarly, yet another existing solution defines a complex junction as a collection of simple junctions, including T-junctions, fork junctions, and crossing junctions, and it is aimed to characterize and detect different forms of complex junctions based on compositional statistics derived from the simple junctions. Yet another existing solution leverages domain knowledge of road and junction design to establish a purely topological definition of complex junctions as groups of nodes representing secondary roads isolated from the rest of the network by relevant nodes of primary roads.

[0016] However, all of these existing solutions primarily focus on geometric and semantic simplification of complex junctions, rather than enhancing the realism of their representation. The existing solutions do not fully account for the wide variations in shape, node density, and connectivity that real-world complex junctions exhibit. Moreover, the representation of junctions is often scale-dependent, and there is currently no unified approach to transition from a single-line representation to a dual-line representation, as required, for instance, in the conversion from an Environmental Systems Research Institute, ESRI, shapefile format to standard traffic simulation software input. In such cases, junctions are typically created by extending the line segments of the rightmost and leftmost road borders, as outlined in the Simulation of Urban Mobility, SUMO, reference manual.

[0017] Additionally, existing solutions based on a statistical analysis of geometric and topological properties of junctions, as well as road segmentation techniques, commonly encounter challenges in determining appropriate threshold settings, which hinders their generalizability and effectiveness.

[0018] Therefore, there arises a need to address the aforementioned technical problem and drawbacks of the known solutions for correcting junction shapes in a road network map. SUMMARY

[0019] It is an object of the disclosure to provide a method for correcting junction shapes in a road network map and an apparatus for correcting junction shapes in the road network map while avoiding one or more disadvantages of prior art approaches.

[0020] This object is achieved by the features of the independent claims. Further, implementation forms are apparent from the dependent claims, the description, and the figures.

[0021] The disclosure provides a method for correcting junction shapes in a road network map and an apparatus for correcting junction shapes in the road network map.

[0022] According to a first aspect, there is provided a method of correcting junction shapes in a road network map. The method includes obtaining crosswalk features by a classifier from an aerial image of a road and an authoritative road network map of the road. The method includes creating an enhanced road network map by co-locating the crosswalk features as crosswalk instances into the authoritative road network map. Each crosswalk instance in the enhanced road network map includes coordinates of a boundary of a single crosswalk. The method includes grouping the crosswalk instances of the enhanced road network map into one or more junction instances by a geometrical reasoner. The method includes refining the coordinates of the boundaries of the crosswalk instances grouped into the junction instances based on coordinate centers of the junctions from the authoritative road network map by the geometrical reasoner. The method includes updating road network nodes of the enhanced road network map by repositioning and / or removing one or more road network nodes of the authoritative road network map and / or adding one or more new road network nodes to the enhanced road network map by the geometrical reasoner. The method includes updating road network segments of the enhanced road network map by one or more of repositioning, resizing, and removing one or more road network segments of the authoritative road network map in accordance with the updated road network nodes by the geometrical reasoner. The method includes updating attributes of the updated road network nodes and the updated road network segments in the enhanced road network map to provide for a connectivity of the enhanced road network by the geometrical reasoner.

[0023] The method directly incorporates corrections obtained from aerial imagery features (e.g., road network features extracted through a DL segmentation algorithm) into the authoritative road network map and automatically enhances the junction shape within the authoritative road network map. The method modifies the geometric properties of the authoritative map data, including road network nodes and / or road network edges positioning, and their associated metadata. By being independent of specific simulator formats, the method ensures accurate alignment of the simulated road network map with real-world conditions, enhancing the overall accuracy of traffic simulations. The method defines a shape and topology of the signalized / compl ex junctions from the road network map retrieved features (e.g., semantic and instance segmentation operated on crosswalk features) and subsequent geo-colocation of the crosswalk instances in the authoritative road network map and alignment (i.e., geospatial reasoning) to the network map properties. Thus, the method can generate a realistic and detailed topological representation of complex junctions within the road network map.

[0024] The method enables the automation of the junction retrieval and correction task by combining the DL segmentation model for crosswalks feature extraction and the geometrical reasoner for structural correction. This combination eliminates the need for manual validation, as the method intelligently applies corrections based on rules or models and the extracted crosswalk features. Additionally, the method implements a validation step to verify the preservation of basic connectivity road rules during the conversion process and provide the connectivity of the enhanced road network. The utilization of distinct and easily recognizable features, such as crosswalks, enables the method to simplify the correction process and employ low-resolution aerial images for junction correction. This is possible because standard algorithms can achieve high precision in identifying and extracting such features without requiring extensive training. Consequently, the method can efficiently make corrections based on these features, thereby reducing the complexity and resource requirements associated with higher-resolution imagery or pixel-level segmentation approaches. The simplicity of extracting and leveraging these crosswalks features enables more efficient and effective use of available aerial / satellite imagery, resulting in a practical and accessible solution for junction corrections.

[0025] The method utilizes a unique pipeline including the DL model, sequential rules and translations to enhance shapes of crosswalks and junctions within the authoritative road network map. This DL model, which includes semantic assignment, position refinement, update of road network nodes and road network edges, and metadata validation, distinguishes itself from traditional heuristic-based approaches. In an implementation, the unique pipeline of corrections applied in the method to the authoritative road network map can be isolated as a separate pre-processing component for enhancing user input map data within a traffic simulation software, leading to more accurate and realistic simulations. The method can also be relevant and beneficial for other industrial and consumer-oriented applications, for example, zebra crossing classification and detection to assist visually impaired people, automated driving navigation systems for pedestrians and vehicles, environment monitoring, and urban planning for municipalities.

[0026] Optionally, the method further includes converting the enhanced road network map into a predefined road network format for a road traffic simulator. Optionally, the refining of the coordinates of the boundaries of the crosswalk instances grouped into the junction instances includes shifting the boundaries of the crosswalk instances grouped into each junction instance with respect to a coordinate center of the junction from the authoritative road network map.

[0027] Optionally, the obtaining of crosswalk features from the aerial image of the road includes retrieving image features relevant for crosswalks by the classifier with a Deep Learning, DL, segmentation algorithm. Optionally, the method further includes training a DL model of the DL segmentation algorithm based on input aerial images of roads with crosswalks, centered on positions of road network nodes associated with a junction of the authoritative road network map and annotated pixel-wise into classes of crosswalks and background components, using transfer learning starting from a pre-trained model with applying different types of data augmentation to the input images.

[0028] Optionally, the obtaining of crosswalk features from the aerial image of the road includes retrieving image features by the classifier with a semantic segmentation method and instantiating single crosswalk objects. Optionally, the method further includes providing map features for the enhanced road network map by a user. The map features include one or more of road network instances of different types, attributes, and natural terrain characteristics. Optionally, the creating of the enhanced road network map includes co-locating the crosswalk features as crosswalk instances into the authoritative road network map using coordinate projection and semantic assignment.

[0029] Optionally, the grouping of the crosswalk instances of the enhanced road network map into the junction instances is based on one or more of semantic information about a road network node on which the input aerial image was centered, a proximity of a crosswalk to a nearest junction, and a junction importance defined by a number of zebra crossings associated with the junction.

[0030] Optionally, the refining of the coordinates of the boundaries of the crosswalk instances grouped into each junction instance comprises: (i) shifting the coordinates of the boundaries of the grouped crosswalk instances all together to move their coordinate center into a coordinate center of the junction according to the authoritative road network map, and (ii) shifting the coordinates of the boundaries of one or more of the grouped crosswalk instances horizontally and / or vertically to center their positions with respect to positions of the road network segments in the enhanced road network map.

[0031] Optionally, the grouping of the crosswalk instances of the enhanced road network map into the junction instances and / or the refining of the coordinates of the boundaries of the crosswalk instances grouped into each junction instance includes using one or more of a machine learning, ML-based model, a distance-based heuristic model, a Neural Network, and a non-maximal suppression algorithm.

[0032] Optionally, the updating of road network nodes of the enhanced road network map includes, for each junction instance: (i) generating a junction boundary to enclose the boundaries of all the crosswalk instances grouped into the junction instance, (ii) determining cut points between edges of the junction boundary and the boundaries of the crosswalk instances grouped into the junction instance using a spatial index, and (iii) one or more of repositioning all road network nodes of the enhanced road network map located within the junction boundary to edges of the junction boundary, removing one or more road network nodes within the junction boundary, and adding one or more new road network nodes outside the junction boundary using the cut points.

[0033] Optionally, the updating of road network nodes of the enhanced road network map further includes determining updated positions of the road network nodes based on an information about one or more of lane numbers for different moving directions of the road, a road and / or lane width, and internal junction markings.

[0034] Optionally, the updating of road network segments of the enhanced road network map includes, for each junction instance: (i) repositioning a starting or an ending point of each road network segment related to the junction instance and connected to one of the road network nodes originally located within the junction boundary, so that the repositioned starting or ending point coincides with an updated position of the road network node at the edges of the junction boundary, wherein parts of the repositioned road network segments originally located within the junction boundary are removed, and (ii) translating each road network segment that connects two road network nodes originally located within the junction boundary to a segment outside of the junction boundary, so that predecessors and successors of the two road network nodes are preserved and a length of the translated road network segment is resized to be proportional to an original length of the road network segment within the junction boundary.

[0035] Optionally, the updating of attributes of the road network nodes and the road network segments in the enhanced road network map includes: (i) updating metadata of the enhanced road network map in accordance with updates of the road network nodes, and (ii) updating metadata of the enhanced road network map in accordance with updates of the road network segments.

[0036] Optionally, the method further includes correcting lane connections in the enhanced road network map with the updated road network nodes and the updated road network segments by one or more of identifying successor and predecessor lanes for a given road network segment, removing overlapping nodes, and determining permitted movement directions at intersections.

[0037] According to a second aspect, there is provided an apparatus for correcting junction shapes in a road network map. The apparatus includes an input module and a geometrical reasoner processing module. The input module is configured for: (i) obtaining crosswalk features with a classifier from an aerial image of a road and an authoritative road network map of the road, and (ii) creating an enhanced road network map by co-locating the crosswalk features as crosswalk instances into the authoritative road network map. Each crosswalk instance in the enhanced road network map includes coordinates of a boundary of a single crosswalk. The geometrical reasoner processing module is configured for: (i) grouping the crosswalk instances of the enhanced road network map into one or more junction instances, (ii) refining the coordinates of the boundaries of the crosswalk instances grouped into the junction instances based on coordinate centers of the junctions from the authoritative road network map, (iii) updating road network nodes of the enhanced road network map by repositioning and / or removing one or more road network nodes of the authoritative road network map and / or adding one or more new road network nodes to the enhanced road network map, (iv) updating road network segments of the enhanced road network map by one or more of repositioning, resizing and removing one or more road network segments of the authoritative road network map in accordance with the updated road network nodes, and (v) updating attributes of the updated road network nodes and the updated road network segments in the enhanced road network map to provide for a connectivity of the enhanced road network.

[0038] The apparatus directly incorporates corrections obtained from aerial imagery features (e.g., road network features extracted through a DL segmentation algorithm) into the authoritative road network map and automatically enhances the junction shape within the authoritative road network map. The apparatus modifies the geometric properties of the authoritative map data, including road network nodes and / or road network edges positioning, and their associated metadata. By being independent of specific simulator formats, the apparatus ensures accurate alignment of the simulated road network map with real-world conditions, enhancing the overall accuracy of traffic simulations. The apparatus defines a shape and topology of the signalized / compl ex junctions from the road network map retrieved features (e.g., semantic and instance segmentation operated on crosswalk features) and subsequent geo-colocation of the crosswalk instances in the authoritative road network map and alignment (i.e., geospatial reasoning) to the network map properties. Thus, the apparatus is able to generate a realistic and detailed topological representation of complex junctions within the road network map.

[0039] The apparatus enables the automation of the junction retrieval and correction task by combining the DL segmentation model for crosswalks feature extraction and the geometrical reasoner processing module for structural correction. This combination eliminates the need for manual validation, as the apparatus intelligently applies corrections based on rules and the extracted crosswalk features. Additionally, the apparatus implements a validation step to verify the preservation of basic connectivity road rules during the conversion process. The utilization of distinct and easily recognizable features, such as crosswalks, enables the apparatus to simplify the correction process and employ low-resolution aerial images for junction correction. This is possible because standard algorithms can achieve high precision in identifying and extracting such features without requiring extensive training. Consequently, the apparatus can efficiently make corrections based on these features, thereby reducing the complexity and resource requirements associated with higher-resolution imagery or pixel-level segmentation approaches. The simplicity of extracting and leveraging these crosswalks features enables more efficient and effective use of available aerial / satellite imagery, resulting in a practical and accessible solution for junction corrections.

[0040] The apparatus utilizes a unique pipeline including the DL model, sequential rules and translations to enhance crosswalks and junctions within the authoritative road network map. This DL model, which includes semantic assignment, position refinement, update of road network nodes and road network edges, and metadata validation, distinguishes itself from traditional heuristic-based approaches. In an implementation, the unique pipeline of corrections applied by the apparatus to the authoritative road network map can be isolated as a separate pre-processing component for enhancing user input map data within a traffic simulation software, leading to more accurate and realistic simulations. The apparatus can also be relevant for other industrial and consumer-oriented applications, for example, zebra crossing classification and detection to assist visually impaired people, automated driving navigation systems for pedestrians and vehicles, environment monitoring, and urban planning for municipalities.

[0041] Optionally, the apparatus further includes an output module that is configured for converting the enhanced road network map into a pre-defined road network format for a road traffic simulator. Optionally, the geometrical reasoner processing module is configured for shifting the boundaries of the crosswalk instances grouped into each junction instance with respect to a coordinate center of the junction from the authoritative road network map. Optionally, the classifier is configured for retrieving image features relevant for crosswalks from the aerial image of the road with a Deep Learning, DL, segmentation algorithm.

[0042] Optionally, the input module is configured for training a DL model of the DL instance segmentation algorithm based on input aerial images of roads with crosswalks, centered on positions of road network nodes associated with a junction of the authoritative road network map and annotated pixel-wise into classes of crosswalks and background components, using transfer learning starting from a pre-trained model with applying different types of data augmentation to the input images.

[0043] Optionally, the classifier is configured for retrieving image features from the aerial image of the road with a semantic segmentation method and for instantiating single crosswalk objects. Optionally, the input module is further configured for obtaining map features for the enhanced road network map provided by a user. The map features include one or more of road network instances of different types, attributes, and natural terrain characteristics. Optionally, the input module is configured for co-locating the crosswalk features as crosswalk instances into the authoritative road network map using coordinate projection and semantic assignment.

[0044] Optionally, the geometrical reasoner processing module is configured for grouping of the crosswalk instances of the enhanced road network map into the junction instances based on one or more of semantic information about a road network node on which the input aerial image was centered, a proximity of a crosswalk to a nearest junction, and a junction importance defined by a number of zebra crossings associated with the junction. Optionally, the geometrical reasoner processing module is configured for refining the coordinates of the boundaries of the crosswalk instances grouped into each junction instance by means of (i) shifting the coordinates of the boundaries of the grouped crosswalk instances all together to move their coordinate center into a coordinate center of the junction according to the authoritative road network map, and (ii) shifting the coordinates of the boundaries of one or more of the grouped crosswalk instances horizontally and / or vertically to center their positions with respect to positions of the road network segments in the enhanced road network map.

[0045] Optionally, the geometrical reasoner processing module is configured for grouping the crosswalk instances of the enhanced road network map into the junction instances and / or refining the coordinates of the boundaries of the crosswalk instances grouped into each junction instance by means of using one or more of a machine learning, ML-based model, a distancebased heuristic model, a Neural Network, and a non-maximal suppression algorithm.

[0046] Optionally, the geometrical reasoner processing module is configured for updating road network nodes of the enhanced road network map, for each junction instance, by means of (i) generating a junction boundary to enclose the boundaries of all the crosswalk instances grouped into the junction instance, (ii) determining cut points between edges of the junction boundary and the boundaries of the crosswalk instances grouped into the junction instance using a spatial index, and (iii) one or more of repositioning all road network nodes of the enhanced road network map located within the junction boundary to edges of the junction boundary, removing one or more road network nodes within the junction boundary, and adding one or more new road network nodes outside the junction boundary using the cut points.

[0047] Optionally, the geometrical reasoner processing module is further configured for determining updated positions of the road network nodes based on an information about one or more of lane numbers for different moving directions of the road, a road and / or lane width, and internal junction markings.

[0048] Optionally, the geometrical reasoner processing module is configured for updating road network segments of the enhanced road network map, for each junction instance, by means of (i) repositioning a starting or an ending point of each road network segment related to the junction instance and connected to one of the road network nodes originally located within the junction boundary, so that the repositioned starting or ending point coincides with an updated position of the road network node at the edges of the junction boundary, wherein parts of the repositioned road network segments originally located within the junction boundary are removed, and (ii) translating each road network segment that connects two road network nodes originally located within the junction boundary to a segment outside of the junction boundary, so that predecessors and successors of the two road network nodes are preserved and a length of the translated road network segment is resized to be proportional to an original length of the road network segment within the junction boundary.

[0049] Optionally, the geometrical reasoner processing module is configured for updating attributes of the road network nodes and the road network segments by means of (i) updating metadata of the enhanced road network map in accordance with updates of the road network nodes, and (ii) updating metadata of the enhanced road network map in accordance with updates of the road network segments.

[0050] Optionally, the geometrical reasoner processing module is further configured for correcting lane connections in the enhanced road network map with the updated road network nodes and the updated road network segments by one or more of identifying successor and predecessor lanes for a given road network segment, removing overlapping nodes and determining permitted movement directions at intersections.

[0051] Therefore, in contradistinction to the existing solutions, the apparatus directly incorporates corrections obtained from aerial imagery features (e.g., road network features extracted through the DL segmentation algorithm) into the authoritative road network map and automatically enhances the junction shape within the authoritative road network map. The apparatus modifies the geometric properties of the authoritative map data, including road network nodes and / or road network edges positioning, and their associated metadata. By being independent of specific simulator formats, the apparatus ensures accurate alignment of the simulated road network map with real-world conditions, enhancing the overall accuracy of traffic simulations. The apparatus defines a shape and topology of the signalized / compl ex junctions from the road network map retrieved features (e.g., semantic and instance segmentation operated on crosswalk features) and subsequent geo-colocation of the crosswalk instances in the authoritative road network map and alignment (i.e., geospatial reasoning) to the network map properties. Thus, the apparatus is able to generate a realistic and detailed topological representation of complex junctions within the road network map. These and other aspects of the disclosure will be apparent from the implementation(s) described below.

[0052] BRIEF DESCRIPTION OF DRAWINGS

[0053] Implementations of the disclosure will now be described, by way of example only, with reference to the accompanying drawings, in which:

[0054] FIG. 1 is a block diagram that illustrates an apparatus for correcting junction shapes in a road network map in accordance with an implementation of the disclosure;

[0055] FIG. 2 is an exemplary illustration of components of an apparatus for correcting junction shapes in a road network map of FIG. 1 and operations realized by the components in accordance with an implementation of the disclosure;

[0056] FIG. 3 is an exemplary illustration of components of an apparatus for correcting junction shapes in a road network map of FIG. 2 and graphical explanations of operations realized by the components in accordance with an implementation of the disclosure;

[0057] FIG. 4 illustrates an exemplary crosswalk feature extraction pipeline in an apparatus for correcting junction shapes in a road network map in accordance with an implementation of the disclosure;

[0058] FIGS. 5A-5B are flow diagrams that illustrate a method of correcting junction shapes in a road network map in accordance with an implementation of the disclosure; and

[0059] FIG. 6 is an illustration of a computer system (e.g., a geometrical reasoner processing module) in which the various architectures and functionalities of the various previous implementations may be implemented.

[0060] DETAILED DESCRIPTION OF THE DRAWINGS

[0061] Implementations of the disclosure provide a method for correcting junction shapes in a road network map and an apparatus for correcting junction shapes in the road network map.

[0062] To make solutions of the disclosure more comprehensible for a person skilled in the art, the following implementations of the disclosure are described with reference to the accompanying drawings. Terms such as "a first", "a second", "a third", and "a fourth" (if any) in the summary, claims, and foregoing accompanying drawings of the disclosure are used to distinguish between similar objects and are not necessarily used to describe a specific sequence or order. It should be understood that the terms so used are interchangeable under appropriate circumstances, so that the implementations of the disclosure described herein are, for example, capable of being implemented in sequences other than the sequences illustrated or described herein. Furthermore, the terms "include" and "have" and any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, a method, a system, a product, or a device that includes a series of steps or units, is not necessarily limited to expressly listed steps or units but may include other steps or units that are not expressly listed or that are inherent to such process, method, product, or device.

[0063] FIG. 1 is a block diagram that illustrates an apparatus 100 for correcting junction shapes in a road network map in accordance with an implementation of the disclosure. The apparatus 100 includes an input module 102 and a geometrical reasoner processing module 104. The input module 102 is configured for: (i) obtaining crosswalk features with a classifier from an aerial image of a road and an authoritative road network map of the road, and (ii) creating an enhanced road network map by co-locating the crosswalk features as crosswalk instances into the authoritative road network map. Each crosswalk instance in the enhanced road network map includes coordinates of a boundary of a single crosswalk. The geometrical reasoner processing module 104 is configured for: (i) grouping the crosswalk instances of the enhanced road network map into one or more junction instances, (ii) refining the coordinates of the boundaries of the crosswalk instances grouped into the junction instances based on coordinate centers of the junctions from the authoritative road network map, (iii) updating road network nodes of the enhanced road network map by repositioning and / or removing one or more road network nodes of the authoritative road network map and / or adding one or more new road network nodes to the enhanced road network map, (iv) updating road network segments of the enhanced road network map by one or more of repositioning, resizing and removing one or more road network segments of the authoritative road network map in accordance with the updated road network nodes, and (v) updating attributes of the updated road network nodes and the updated road network segments in the enhanced road network map to provide for a connectivity of the enhanced road network.

[0064] The apparatus 100 directly incorporates corrections obtained from aerial imagery features (e.g., road network features extracted through a DL segmentation algorithm) into the authoritative road network map and automatically enhances the junction shape within the authoritative road network map. The apparatus 100 modifies the geometric properties of the authoritative map data, including road network nodes and / or road network edges positioning, and their associated metadata. By being independent of specific simulator formats, the apparatus 100 ensures accurate alignment of the simulated road network map with real-world conditions, enhancing the overall accuracy of traffic simulations. The apparatus 100 defines a shape and topology of the signalized / complex junctions from the road network map retrieved features (e.g., semantic and instance segmentation operated on crosswalk features) and subsequent geo-colocation of the crosswalk instances in the authoritative road network map and alignment (i.e., geospatial reasoning) to the network map properties. Thus, the apparatus 100 can generate a realistic and detailed topological representation of complex junctions within the road network map.

[0065] The apparatus 100 enables the automation of the junction retrieval and correction task by combining the DL segmentation model for crosswalks feature extraction and the geometrical reasoner processing module 104 for structural correction. This combination eliminates the need for manual validation, as the apparatus 100 intelligently applies corrections based on rules and the extracted crosswalk features. Additionally, the apparatus 100 implements a validation step to verify the preservation of basic connectivity road rules during the conversion process. The utilization of distinct and easily recognizable features, such as crosswalks, enables the apparatus 100 to simplify the correction process and employ low-resolution aerial images for junction correction. This is possible because standard algorithms can achieve high precision in identifying and extracting such features without requiring extensive training. Consequently, the apparatus 100 can efficiently make corrections based on these features, thereby reducing the complexity and resource requirements associated with higher-resolution imagery or pixel-level segmentation approaches. The simplicity of extracting and leveraging these crosswalks features enables more efficient and effective use of available aerial / satellite imagery, resulting in a practical and accessible solution for junction corrections.

[0066] The apparatus 100 utilizes a unique DL model of sequential rules and translations to enhance crosswalks and junctions within the authoritative road network map. This DL model, which includes semantic assignment, position refinement, road network nodes, road network edges repositioning, and metadata validation, distinguishes itself from traditional heuristic-based approaches. By isolating the DL model as a separate pre-processing component, it becomes a specialized tool for enhancing user input map data within a traffic simulation, leading to more accurate and realistic simulations. The apparatus 100 can also be relevant for other industrial and consumer-oriented applications, for example, zebra crossing classification and detection to assist visually impaired people, automated driving navigation systems for pedestrians and vehicles, environment monitoring, and urban planning for municipalities.

[0067] Optionally, the input module 102 includes features extracted from the aerial image, represented as polygons, and the authoritative road network map of the road in the form of an Environmental Systems Research Institute, ESRI, shapefile. The features of the aerial image and the authoritative road network map may be obtained using a Deep Learning, DL, segmentation algorithm. Optionally, the classifier is configured for retrieving image features relevant for crosswalks from the aerial image of the road with the Deep Learning, DL, segmentation algorithm. Optionally, the classifier is configured for retrieving image features from the aerial image of the road with a semantic segmentation method and for instantiating single crosswalk objects. Optionally, the crosswalk features are co-located on the authoritative road network map to create the enhanced road network map. Optionally, the input module 102 is configured for co-locating the crosswalk features as crosswalk instances into the authoritative road network map using coordinate projection and semantic assignment. Optionally, the input module 102 is further configured for obtaining map features for the enhanced road network map provided by a user. Alternatively, the map features for the enhanced road network map can be obtained from the DL model configured to operate as a separate pre-processing component of a traffic simulation software. The map features include one or more of road network instances of different types, attributes, and natural terrain characteristics. Optionally, the authoritative road network map / shapefile includes the essential components such as the geometric graph of the geographic elements (.shp file), the geometry indexes (.shx file), and the feature attributes (*.dbf). This file format represents the road network and junctions as a collection of road network edges and connecting road network nodes, accompanied by associated attributes.

[0068] The geometrical reasoner processing module 104 serves as a central system that performs all automated functionalities, whether rule-based or model-based, for correcting the authoritative road network map. These functionalities encompass tasks such as assigning semantic labels to crosswalks, centering, shifting, and modifying relevant geometric elements (i.e., road network nodes and road network edges) within the authoritative road network map / shapefile, and ensuring metadata consistency to preserve map connectivity. Optionally, the repositioning of road network nodes and road network edges goes beyond mere adjustments of generic map attributes. Instead, the geometrical reasoner processing module 104 directly modifies the structural elements of the road network itself, such as the lengths of road network edges, coordinates of starting and ending points, and coordinates of road network nodes. Optionally, the geometrical reasoner processing module 104 incorporates a geometry-aware machine learning, ML, or deep learning, DL to enhance a wide range of semantic, spatial, and multilevel models executed in a sequential manner. Although the input and output of the various components are interdependent, the validation of these models can be conducted on a component-wise basis.

[0069] Optionally, the apparatus 100 further includes an output module 106 that is configured for converting the enhanced road network map into a pre-defined road network format for a road traffic simulator. For instance, if the objective of the apparatus 100 is to achieve a more precise representation of junctions for a road traffic simulation, the output module 106 may convert the enhanced road network map into a road network representation compatible with the targeted road traffic simulator (i.e., specific use-case). During this step, the line-lane road representation may be transformed into a dual-lane (or even 3 -dimensional) representation, typically following a predefined set of rules. The modifications made to the authoritative road network map may be reflected in the resulting output (i.e., pre-defined road network format), resulting in more realistic junction shapes and dimensions. This ensures that the structural changes made to the authoritative road network map are accurately translated into a representation that aligns with the desired simulation requirements.

[0070] FIG. 2 is an exemplary illustration of components of an apparatus 200 for correcting junction shapes in a road network map, such as the apparatus 100 of FIG. 1, and operations realized by the components in accordance with an implementation of the disclosure. The apparatus 200 includes an input module 202, a geometrical reasoner processing module 204, and an output module 206. At an operation 214, the input module 202 obtains crosswalk features with a classifier 212 from an aerial image 208 of a road and an authoritative road network map / shapefile 210 of the road and creates an enhanced road network map by co-locating the crosswalk features as crosswalk instances into the authoritative road network map / shapefile 210. Each crosswalk instance in the enhanced road network map includes coordinates of a boundary of a single crosswalk. Optionally, the crosswalk instance includes a pixel-wise map of the crosswalks and classes probabilities, metadata (e.g., a junction the crosswalk belongs to), and background components. Optionally, the input module 202 is configured for co-locating the crosswalk features as crosswalk instances into the authoritative road network map 210 using coordinate projection and semantic assignment. Optionally, the input module 202 is further configured for obtaining map features for the enhanced road network map provided by a user. The map features include one or more of road network instances of different types, attributes, and natural terrain characteristics. Optionally, the classifier 212 is configured for retrieving image features relevant for crosswalks from the aerial image 208 of the road with a Deep Learning, DL, segmentation algorithm. Optionally, the classifier 212 is configured for retrieving image features from the aerial image 208 of the road with a semantic segmentation method and for instantiating single crosswalk objects. Optionally, the input module 202 is configured for training a DL model of the DL segmentation algorithm based on input aerial images of roads with crosswalks, centered on positions of road network nodes associated with a junction of the authoritative road network map 210 and annotated pixel-wise into classes of crosswalks and background components, using transfer learning starting from a pre-trained model with applying different types of data augmentation to the input images.

[0071] At an operation 216, the geometrical reasoner processing module 204 groups / assigns the crosswalk instances of the enhanced road network map into one or more corresponding junction instances. Optionally, the geometrical reasoner processing module 204 is configured for shifting the boundaries of the crosswalk instances grouped into each junction instance with respect to a coordinate center of the junction from the authoritative road network map 210. Optionally, the geometrical reasoner processing module 204 is configured for grouping of the crosswalk instances of the enhanced road network map into the junction instances based on one or more of semantic information about a road network node on which the input aerial image was centered, a proximity of a crosswalk to a nearest junction, and a junction importance defined by a number of zebra crossings associated with the junction.

[0072] At an operation 218, the geometrical reasoner processing module 204 refines the coordinates of the boundaries of the crosswalk instances grouped into the junction instances based on coordinate centers of the junctions from the authoritative road network map 210. Optionally, the geometrical reasoner processing module 204 is configured for refining the coordinates of the boundaries of the crosswalk instances grouped into each junction instance by means of (i) shifting the coordinates of the boundaries of the grouped crosswalk instances all together to move their coordinate center into a coordinate center of the junction according to the authoritative road network map 210, and (ii) shifting the coordinates of the boundaries of one or more of the grouped crosswalk instances horizontally and / or vertically to center their positions with respect to positions of the road network segments in the enhanced road network map. Optionally, the geometrical reasoner processing module 204 is configured for grouping the crosswalk instances of the enhanced road network map into the junction instances and / or refining the coordinates of the boundaries of the crosswalk instances grouped into each junction instance by means of one or more of a machine learning, ML-based model, a distance-based heuristic model, a Neural Network, and a non-maximal suppression algorithm.

[0073] At an operation 220, the geometrical reasoner processing module 204 updates road network nodes of the enhanced road network map by repositioning and / or removing one or more road network nodes of the authoritative road network map 210 and / or adding one or more new road network nodes to the enhanced road network map. Optionally, the geometrical reasoner processing module 204 is configured for updating road network nodes of the enhanced road network map, for each junction instance, by means of (i) generating a junction boundary to enclose the boundaries of all the crosswalk instances grouped into the junction instance, (ii) determining cut points between edges of the junction boundary and the boundaries of the crosswalk instances grouped into the junction instance using a spatial index, and (iii) one or more of repositioning all road network nodes of the enhanced road network map located within the junction boundary to edges of the junction boundary, removing one or more road network nodes within the junction boundary, and adding one or more new road network nodes outside the junction boundary using the cut points. Optionally, the geometrical reasoner processing module 204 is further configured for determining updated positions of the road network nodes based on an information about one or more of lane numbers for different moving directions of the road, a road and / or lane width, and internal junction markings. At a step 222, the geometrical reasoner processing module 204 updates road network segments / edges of the enhanced road network map by one or more of repositioning, resizing, and removing one or more road network segments of the authoritative road network map 210 in accordance with the updated road network nodes. Optionally, the geometrical reasoner processing module 204 is configured for updating road network segments of the enhanced road network map, for each junction instance, by means of (i) repositioning a starting or an ending point of each road network segment related to the junction instance and connected to one of the road network nodes originally located within the junction boundary, so that the repositioned starting or ending point coincides with an updated position of the road network node at the edges of the junction boundary, where parts of the repositioned road network segments originally located within the junction boundary are removed, and (ii) translating each road network segment that connects two road network nodes originally located within the junction boundary to a segment outside of the junction boundary, so that predecessors and successors of the two road network nodes are preserved and a length of the translated road network segment is resized to be proportional to an original length of the road network segment within the junction boundary.

[0074] At an operation 224, the geometrical reasoner processing module 204 updates attributes / connections of the updated road network nodes and the updated road network segments in the enhanced road network map to provide for connectivity of the enhanced road network, which is vital to guarantee full functionality and accuracy of road traffic simulations. Optionally, the geometrical reasoner processing module 204 is configured for updating attributes of the road network nodes and the road network segments by means of (i) updating metadata of the enhanced road network map in accordance with updates of the road network nodes, and (ii) updating metadata of the enhanced road network map in accordance with updates of the road network segments. Optionally, the geometrical reasoner processing module 204 is further configured for correcting lane connections in the enhanced road network map with the updated road network nodes and the updated road network segments by one or more of identifying successor and predecessor lanes for a given road network segment, removing overlapping nodes and determining permitted movement directions at intersections. At an operation 226, the geometrical reasoner processing module 204 obtains the corrected road network map (i.e., the map corrected as compared with the authoritative road network map 210) that represents a final version of the enhanced road network map. Then, at an operation 228, the output module 206 converts the enhanced road network map (i.e., the corrected map) into a pre-defined / simulator-specific road network format 230 for a road traffic simulator through a standardized conversion process.

[0075] FIG. 3 is an exemplary illustration of components of an apparatus 300 for correcting junction shapes in a road network map, such as the apparatus 200 of FIG. 2, and graphical explanations of operations realized by the components in accordance with an implementation of the disclosure. The apparatus 300 includes an input module 302, a geometrical reasoner processing module 304 and an output module 306. The road network maps / images with crosswalks polygons may be derived from a traffic simulation software such as the graphical City Mobility Simulator, CityMoS, Suite editor.

[0076] As illustrated in FIG. 2 with reference to the classifier 212, a segmentation model such as a Deep Learning, DL, segmentation algorithm (e.g., a Mask Region-Based Convolutional Neural Networks, Mask R-CNN, model) is employed in the input module 302 to extract image / crosswalk features relevant to road network correction from an aerial image / satellite image 308 of a road and an authoritative road network map / shapefile 310 of the road. The retrieved crosswalk features are co-located by the input module 302 as crosswalk instances 312 that for illustration purposes are superimposed in FIG. 3 on both the aerial image / satellite image 308 and the authoritative road network map / shapefile 310. A user, through the input module 302 is enabled to create and label aerial image tiles, initiate training a deep learning, DL, model which is the DL segmentation algorithm, and inference scripts for processing new satellite data, which enables functionality for post-processing the results, and aligns the features with a road traffic simulator network map. Optionally, the input module 302 trains the DL model of the DL segmentation algorithm based on input aerial images of roads with crosswalks, centered on positions of road network nodes associated with a junction of the authoritative road network map 310 and annotated pixel-wise into classes of crosswalks and background components, using transfer learning starting from a pre-trained model (e.g., the Mask R-CNN, model trained on a widely known Common Objects in Context, COCO dataset) with applying different types of data augmentation to the input images. Optionally, a large-scale image recognition dataset, such as the COCO dataset, is used for object detection, segmentation, and captioning tasks. The input module 302 creates an enhanced road network map 314 by co-locating the crosswalk features as crosswalk instances 312 into the authoritative road network map 310. Optionally, the input module 302 conducts the co-location of individual crosswalk instances 312, obtained through the DL segmentation algorithm, with the authoritative road network map / shapefile 310 using coordinate projection and semantic assignment. Each crosswalk instance 312 in the enhanced road network map includes coordinates of a boundary 316 of a single crosswalk. In an implementation, the pre-trained model used for inference has been developed using a limited number of manually labeled maps from the city of Xuchang, China.

[0077] The advantages of the use of the transfer learning for training the DL model / segmentation algorithm starting from the pre-trained model include robustness despite limited availability of training data, i.e., by leveraging the transfer learning with the pre-trained model, the DL model can reach higher performance levels, resulting in more accurate outputs even when the apparatus 300 works with a smaller amount of labeled data. The transfer learning provides an additional advantage of reducing training time and conserving computational resources. The transfer learning provides quality of input images and generalization, i.e., the DL model is designed to be highly versatile, and capable of adapting to extract various features visible in the aerial / satellite images. The transfer learning facilitates both binary and multiclass object classification, providing flexibility in the identification of different types of objects. The transfer learning provides post-processing for simplified polygon extraction, i.e., the input images undergo post-processing to derive simplified polygons. The DL model produces single target instances (i.e., crosswalk instances) as model output, making it easier to identify and analyze individual instances within the aerial image / satellite image 308.

[0078] The geometrical reasoner processing module 304 may be divided into a detection block (configured for crosswalk instances 312 semantic assignment and positioning) and a geometric adjustment block (configured for the authoritative road network map / shapefile 310 corrections and validation), though such division is optional and can be of logical nature without physical division of the geometrical reasoner processing module 304 into separate components. Initial operation of the geometrical reasoner processing module 304 is illustrated by a diagram 314 and involves the geometrical reasoner processing module 304 grouping / assigning the crosswalk instances 312 of the enhanced road network map 314, as obtained with the DL segmentation algorithm in the input module 302, into one or more corresponding junction instances. The diagrams 314 shows (on top) a schematic view of a part of the enhanced road network map with a set of crosswalk instances 312 grouped into one junction instance. Furthermore, on bottom of the diagram 314 shown is a schematic view of corresponding enlarged aerial image with some of the crosswalk instances 312 of said junction instance and their boundaries 316. Optionally, the geometrical reasoner processing module 304 is configured for grouping of the crosswalk instances 312 of the enhanced road network map into the junction instances based on one or more of semantic information about a road network node on which the input aerial / satellite image 308 was centered, a proximity of a crosswalk to a nearest junction, and a junction importance defined by a number of zebra crossings associated with the junction. These criteria help in determining which crosswalk instances 312 may be grouped together based on their spatial relationship with nearby junctions. By considering both proximity and junction importance, the grouping process ensures a more comprehensive and accurate representation of the crosswalk instances 312 on the enhanced road network map. Optionally, the grouped crosswalk instances 312 are co-located as the junction instances within a traffic simulator network map, e.g. the CityMoS simulator network map, as in this case shown by the enhanced road network map in the top part of the diagram 314. The traffic simulator network map in this case presents an illustration derived from the traffic simulator user interface, highlighting the co-location of the crosswalks instances 312 grouped into the junction instance.

[0079] As further illustrated by a two-step diagram 318, the geometrical reasoner processing module 304 refines the coordinates of the boundaries 316 of the crosswalk instances 312 grouped into the junction instances based on coordinate centers of the junctions from the authoritative road network map 310, the refined boundaries of the crosswalk instances are shown in a schematic view 320 of a part of the enhanced road network map, one of the refinements being indicated by an arrow. The authoritative road network map / shapefile 310 may be derived from a google satellite map or any other known provider of authoritative road network maps. Optionally, the geometrical reasoner processing module 304 is configured for shifting the boundaries of the crosswalk instances 312 grouped into each junction instance with respect to a coordinate center of the junction from the authoritative road network map 310. In an implementation, the geometrical reasoner processing module 304 is configured for refining the coordinates of the boundaries of the crosswalk instances 312 grouped into each junction instance by means of (i) shifting the coordinates of the boundaries 316 of the grouped crosswalk instances 312 all together to move their coordinate center into a coordinate center of the junction according to the authoritative road network map 310, and (ii) shifting the coordinates of the boundaries 316 of one or more of the grouped crosswalk instances 312 horizontally and / or vertically to center their positions with respect to positions of the road network segments / edges in the enhanced road network map 320, for example, as it is shown in a two-step diagram 322 illustrating the positions of three grouped crosswalk instances before (on top) and after (on bottom) the shifting of their coordinates.

[0080] The presence of inaccuracies in the latitudinal and longitudinal positions of the road network segments / edges and road network nodes in the authoritative road network map / shapefile 310 may lead to significant discrepancies when attempting to overlay the crosswalk instances 312 directly onto the authoritative road network map / shapefile 310. Consequently, the apparatus 300 performs the refining and shifting operations as described above that are carried out to adjust the groups of crosswalk instances 312 in relation to both the original junction group center and the center computed from the coordinates of the corresponding aerial image, such as the aerial image on top of the two-step diagram 318. The advantage of using the crosswalk instances 312 groups is that the relative distances between crosswalk instances 312 and thus the junction topology, are preserved. To achieve proper alignment, the crosswalk instances 312 may be shifted both horizontally and vertically, for example, as it is shown in the two-step diagram 322. This realignment is necessary to center their positions with respect to the corresponding segments / edges in the authoritative road network map / shapefile 310.

[0081] Subsequent operations of correcting junction shapes in the authoritative road network map 310 by the geometrical reasoner processing module 304 involves updating the road network nodes and the road network edges, and the connectivity validation. Regarding the road network nodes, a spatial index is used to calculate the cut points where the road network edges intersect with the boundaries of the crosswalk instances 312. Any internal road network nodes, i.e. ones arranged inside the boundaries of the crosswalk instances grouped into each junction instance, are then updated by means of one or more of repositioning, removing the internal road network nodes and adding new road network nodes so to arrange the internal road network nodes to the external road network edges of the boundaries encompassing the crosswalk instance 312 groups, i.e. the junction instances. The geometrical reasoner processing module 304 updates the road network nodes of the enhanced road network map by repositioning and / or removing one or more road network nodes of the authoritative road network map 310 and / or adding one or more new road network nodes to the enhanced road network map. Optionally, the geometrical reasoner processing module 304 is configured for updating road network nodes of the enhanced road network map, for each junction instance, by means of (i) generating a junction boundary to enclose the boundaries of all the crosswalk instances 312 grouped into the junction instance, (ii) determining cut points between edges of the junction boundary and the boundaries of the crosswalk instances grouped into the junction instance using a spatial index, and (iii) one or more of repositioning all road network nodes of the enhanced road network map located within the junction boundary to edges of the junction boundary, removing one or more road network nodes within the junction boundary, and adding one or more new road network nodes outside the junction boundary using the cut points, for example, as it is shown in a two- step diagram 324 illustrating the positions of the road network nodes before and after the nodes updating operation described above, at that the road network nodes subject to the update are shown as circles in a left part (before the updating) of the diagram 324, and the updated position of one of said nodes is indicated by an arrow 324A.

[0082] To reposition the road network edges in the authoritative road network map / shapefile 310, the apparatus 300 employs a distinct heuristic approach based on the location of each crosswalk instance 312. The positioning of the road network edges is closely intertwined with the placement of road network nodes. In the simplest scenario, where there are no additional road network nodes within the boundary of the grouped crosswalks, i.e. the junction boundary, only the starting or ending points of the road network edges need to be adjusted to align with the new road network node coordinates. At further operation, the geometrical reasoner processing module 304 updates road network segments / edges of the enhanced road network map 320 by one or more of repositioning, resizing, and removing one or more road network segments of the authoritative road network map 310 in accordance with the updated road network nodes, for example, as it is also shown in the two-step diagram 324 illustrating the repositioning, the resizing and the removing of the road network segments / edges after the edges updating operation following to the nodes updating operation, wherein the road network edges subject to the update are segments in-between the circles (nodes) in the left part (before the updating) of the diagram 324 as well as segments inside the boundary of the junction, and one of the updates of the edges is indicated by an arrow 324B. Optionally, the geometrical reasoner processing module 304 is configured for updating road network segments of the enhanced road network map, for each junction instance, by means of (i) repositioning a starting or an ending point of each road network segment related to the junction instance and connected to one of the road network nodes originally located within the junction boundary, so that the repositioned starting or ending point coincides with an updated position of the road network node at the edges of the junction boundary, where parts of the repositioned road network segments originally located within the junction boundary are removed, and (ii) translating each road network segment that connects two road network nodes originally located within the junction boundary to a segment outside of the junction boundary, so that predecessors and successors of the two road network nodes are preserved and a length of the translated road network segment is resized to be proportional to an original length of the road network segment within the junction boundary.

[0083] The next operation of the geometrical reasoner processing module 304 is the metadata consistency validation, wherein the geometrical reasoner processing module 304 updates attributes / metadata of the updated road network nodes and the updated road network segments in the enhanced road network map 320 to provide a connectivity of the enhanced road network, which is vital for functional and accurate road traffic simulations. Optionally, the geometrical reasoner processing module 304 is configured for updating attributes of the road network nodes and the road network segments by means of (i) updating metadata of the enhanced road network map in accordance with updates of the road network nodes, and (ii) updating metadata of the enhanced road network map in accordance with updates of the road network segments. Optionally, the geometrical reasoner processing module 304 is further configured for correcting lane connections in the enhanced road network map with the updated road network nodes and the updated road network segments by one or more of identifying successor and predecessor lanes for a given road network segment, removing overlapping nodes and determining permitted movement directions at intersections. The apparatus 300 performs the attributes / metadata consistency validation operation to guarantee the coherence of attribute information associated with the geometrical elements (e.g., the road network nodes and the road network segments / edges). This operation encompasses various aspects, such as verifying the accuracy of details regarding the starting and ending nodes of the edges. It also ensures the uniqueness of node identifiers, such as the main node designation for a signalized intersection, along with its association with the corresponding junction. By performing this attributes / metadata consistency check, the integrity of the attribute information is maintained, thereby preventing any inconsistencies or discrepancies that may have arisen during the abovedescribed authoritative road network map / shapefile 310 correction process. The attributes / metadata consistency validation operation serves to validate the correctness of the metadata associated with the geometrical elements, thereby ensuring the overall quality and accuracy of the corrected authoritative road network map / shapefile, i.e. the enhanced road network map 320.

[0084] At a further operation illustrated by a two-step diagram 326, the output module 306 converts the enhanced road network map 320 into a pre-defined / simulator-specific road network format, for example, as illustrated by a schematic view on bottom of the diagram 326, for a road traffic simulator through a standardized conversion process. This conversion enables the utilization of the road network within the selected simulator for further analysis and simulation.

[0085] FIG. 4 illustrates an exemplary crosswalk feature extraction pipeline in an apparatus 400 for correcting junction shapes in a road network map in accordance with an implementation of the disclosure. With reference to FIG. 4 shown is an exemplary pipeline 402-404 for extraction of network and map features implemented in the apparatus 400 that employs a Deep Learning, DL, segmentation algorithm. The pipeline as shown in FIG. 4 is employed in the apparatus 400 in accordance with an implementation of the disclosure to extract crosswalk features from aerial / satellite (e.g. google map) images and an authoritative road network map / shapefile of a road relevant to road network correction, specifically the crosswalk instances. The feature extraction pipeline includes (i) tools for the creation of image tiles using aerial (e.g. satellite / google map) images and an authoritative road network map shapefile and their subsequent manual labeling (for example, the manual labeling can be done using proper software, such as the Label Studio), (ii) means for training a DL model of the DL segmentation algorithm with the use of transfer learning starting from the Mask R-CNN model trained on the well-known COCO dataset, (iii) scripts designed for performing inference on newly acquired satellite data, thereby allowing the DL model to make predictions and identify features of interest, and (iv) means for post-processing of the results and co-location of the identified features on the authoritative road network map / shapefile.

[0086] Optionally, the apparatus 400 trains the DL model of the DL segmentation algorithm based on input aerial images of roads with crosswalks, centered on positions of road network nodes associated with a junction of the authoritative road network map and annotated pixel-wise into classes of crosswalks and background components, using transfer learning starting from a pretrained model (e.g., the Mask R-CNN, model trained on the COCO dataset) with applying different types of data augmentation to the input images.

[0087] The advantages of the transfer learning starting from the pre-trained model include availability of training data, i.e., by leveraging the transfer learning with the pre-trained model, the DL model can reach higher performance levels, resulting in more accurate outputs even when the apparatus works with a smaller amount of labeled data. The transfer learning provides an additional advantage of reducing training time and conserving computational resources. The transfer learning provides quality of input images and generalization, i.e., the DL model is designed to be highly versatile, and capable of adapting to extract various features visible in the aerial / satellite images. The transfer learning facilitates both binary and multiclass object classification, providing flexibility in the identification of different types of objects. The transfer learning provides post-processing for simplified polygon extraction, i.e., the input images undergo post-processing to derive simplified polygons. The DL model produces single target instances (i.e., crosswalk instances) as model output, making it easier to identify and analyze individual instances within the aerial image / satellite image.

[0088] Following to the crosswalk feature extraction as described above, the enhanced road network map is created by a geometric reasoner and then converted into a pre-defined road network format for a road traffic simulator by an output module within the apparatus 400 for correcting junction shapes in a road network map.

[0089] The output module of the apparatus 400 for correcting junction shapes in a road network map, such as the output module 306 of FIG. 3, converts the enhanced road network map / corrected authoritative road network map into a pre-defined / simulator-specific road network format for a road traffic simulator through a standardized conversion process. This conversion enables the utilization of the enhanced road network within the selected simulator for further analysis and simulation. An efficiency of the use of an implementation of the present disclosure can be estimated automatically, e.g. using a polygon convex hull enclosing crosswalks groups as a proxy for a reference case, or by manually drawing intersection borders in a real-life road network maps. Such estimation of the efficiency of the correction of junction shapes performed on a large- scale scenario comprising 241 junctions of the real-life road network map of the city of Xuchang, China has shown that the use of the apparatus or the method of the present disclosure improves an overlap between the real junction shape and the junction shape originally obtained from the authoritative road network map / shapefile by a factor of 10 (Intersection over Union before and after the correction process of the present disclosure).

[0090] FIGS. 5A-5B are flow diagrams that illustrate a method of correcting junction shapes in a road network map in accordance with an implementation of the disclosure. At a step 502, crosswalk features are obtained by a classifier from an aerial image of a road and an authoritative road network map of the road. At a step 504, an enhanced road network map is created by co-locating the crosswalk features as crosswalk instances into the authoritative road network map. Each crosswalk instance in the enhanced road network map includes coordinates of a boundary of a single crosswalk. At a step 506, the crosswalk instances of the enhanced road network map are grouped into one or more junction instances by a geometrical reasoner. At a step 508, the coordinates of the boundaries of the crosswalk instances grouped into the junction instances are refined based on coordinate centers of the junctions from the authoritative road network map by the geometrical reasoner. At a step 510, road network nodes of the enhanced road network map are updated by repositioning and / or removing one or more road network nodes of the authoritative road network map and / or adding one or more new road network nodes to the enhanced road network map by the geometrical reasoner. At a step 512, road network segments of the enhanced road network map are updated by one or more of repositioning, resizing, and removing one or more road network segments of the authoritative road network map in accordance with the updated road network nodes by the geometrical reasoner. At a step 514, attributes of the updated road network nodes and the updated road network segments in the enhanced road network map is updated to provide for a connectivity of the enhanced road network by the geometrical reasoner.

[0091] The method directly incorporates corrections obtained from aerial imagery features (e.g., road network features extracted through a DL segmentation algorithm) into the authoritative road network map and automatically enhances the junction shape within the authoritative road network map. The method modifies the geometric properties of the authoritative map data, including road network nodes and / or road network edges positioning, and their associated metadata. By being independent of specific simulator formats, the method ensures accurate alignment of the simulated road network map with real-world conditions, enhancing the overall accuracy of traffic simulations. The method defines a shape and topology of the signalized / compl ex junctions from the road network map retrieved features (e.g., semantic and instance segmentation operated on crosswalk features) and subsequent geo-colocation of the crosswalk instances in the authoritative road network map and alignment (i.e., geospatial reasoning) to the network map properties. Thus, the method is able to generate a realistic and detailed topological representation of complex junctions within the road network map.

[0092] The method enables the automation of the junction retrieval and correction task by combining the DL segmentation model for crosswalks feature extraction and the geometrical reasoner for structural correction. This combination eliminates the need for manual validation, as the method intelligently applies corrections based on rules and the extracted crosswalk features. Additionally, the method implements a validation step to verify the preservation of basic connectivity road rules during conversion process. The utilization of distinct and easily recognizable features, such as crosswalks, enables the method to simplify the correction process and employ low-resolution aerial images for junction correction. This is possible because standard algorithms can achieve high precision in identifying and extracting such features without requiring extensive training. Consequently, the method can efficiently make corrections based on these features, thereby reducing the complexity and resource requirements associated with higher-resolution imagery or pixel-level segmentation approaches. The simplicity of extracting and leveraging these crosswalks features enables more efficient and effective use of available aerial / satellite imagery, resulting in a practical and accessible solution for junction corrections.

[0093] The method utilizes a unique DL model of sequential rules and translations to enhance crosswalks and junctions within the authoritative road network map. This DL model, which includes semantic assignment, position refinement, road network nodes and road network edges repositioning, and metadata validation, distinguishes itself from traditional heuristic-based approaches. By isolating the DL model as a separate pre-processing component, it becomes a specialized tool for enhancing user input map data within a traffic simulation, leading to more accurate and realistic simulations. The method can also be relevant for other industrial and consumer-oriented applications, for example, zebra crossing classification and detection to assist visually impaired people, automated driving navigation systems for pedestrians and vehicles, environment monitoring and urban planning for municipalities.

[0094] Optionally, the method further includes converting the enhanced road network map into a predefined road network format for a road traffic simulator.

[0095] Optionally, the refining of the coordinates of the boundaries of the crosswalk instances grouped into the junction instances includes shifting the boundaries of the crosswalk instances grouped into each junction instance with respect to a coordinate center of the junction from the authoritative road network map.

[0096] Optionally, the obtaining of crosswalk features from the aerial image of the road includes retrieving image features relevant for crosswalks by the classifier with a Deep Learning, DL, segmentation algorithm.

[0097] Optionally, the method further includes training a DL model of the DL segmentation algorithm based on input aerial images of roads with crosswalks, centered on positions of road network nodes associated with a junction of the authoritative road network map and annotated pixelwise into classes of crosswalks and background components, using transfer learning starting from a pre-trained model with applying different types of data augmentation to the input images.

[0098] Optionally, the obtaining of crosswalk features from the aerial image of the road includes retrieving image features by the classifier with a semantic segmentation method and instantiating single crosswalk objects.

[0099] Optionally, the method further includes providing map features for the enhanced road network map by a user. The map features include one or more of road network instances of different types, attributes, and natural terrain characteristics.

[0100] Optionally, the creating of the enhanced road network map includes co-locating the crosswalk features as crosswalk instances into the authoritative road network map using coordinate projection and semantic assignment.

[0101] Optionally, the grouping of the crosswalk instances of the enhanced road network map into the junction instances is based on one or more of semantic information about a road network node on which the input aerial image was centered, a proximity of a crosswalk to a nearest junction, and a junction importance defined by a number of zebra crossings associated with the junction.

[0102] Optionally, the refining of the coordinates of the boundaries of the crosswalk instances grouped into each junction instance comprises: (i) shifting the coordinates of the boundaries of the grouped crosswalk instances all together to move their coordinate center into a coordinate center of the junction according to the authoritative road network map, and (ii) shifting the coordinates of the boundaries of one or more of the grouped crosswalk instances horizontally and / or vertically to center their positions with respect to positions of the road network segments in the enhanced road network map.

[0103] Optionally, the grouping of the crosswalk instances of the enhanced road network map into the junction instances and / or the refining of the coordinates of the boundaries of the crosswalk instances grouped into each junction instance includes one or more of a machine learning, ML- based model, a distance-based heuristic model, a Neural Network, and a non-maximal suppression algorithm.

[0104] Optionally, the updating of road network nodes of the enhanced road network map includes, for each junction instance: (i) generating a junction boundary to enclose the boundaries of all the crosswalk instances grouped into the junction instance, (ii) determining cut points between edges of the junction boundary and the boundaries of the crosswalk instances grouped into the junction instance using a spatial index, and (iii) one or more of repositioning all road network nodes of the enhanced road network map located within the junction boundary to edges of the junction boundary, removing one or more road network nodes within the junction boundary, and adding one or more new road network nodes outside the junction boundary using the cut points.

[0105] Optionally, the updating of road network nodes of the enhanced road network map further includes determining updated positions of the road network nodes based on an information about one or more of lane numbers for different moving directions of the road, a road and / or lane width, and internal junction markings.

[0106] Optionally, the updating of road network segments of the enhanced road network map includes, for each junction instance: (i) repositioning a starting or an ending point of each road network segment related to the junction instance and connected to one of the road network nodes originally located within the junction boundary, so that the repositioned starting or ending point coincides with an updated position of the road network node at the edges of the junction boundary, where parts of the repositioned road network segments originally located within the junction boundary are removed, and (ii) translating each road network segment that connects two road network nodes originally located within the junction boundary to a segment outside of the junction boundary, so that predecessors and successors of the two road network nodes are preserved and a length of the translated road network segment is resized to be proportional to an original length of the road network segment within the junction boundary.

[0107] Optionally, the updating of attributes of the road network nodes and the road network segments in the enhanced road network map includes: (i) updating metadata of the enhanced road network map in accordance with updates of the road network nodes, and (ii) updating metadata of the enhanced road network map in accordance with updates of the road network segments.

[0108] Optionally, the method further includes correcting lane connections in the enhanced road network map with the updated road network nodes and the updated road network segments by one or more of identifying successor and predecessor lanes for a given road network segment, removing overlapping nodes and determining permitted movement directions at intersections.

[0109] FIG. 6 is an illustration of a computer system, e.g., a geometrical reasoner processing module, in which the various architectures and functionalities of the various previous implementations may be implemented. As shown, the computer system 600 includes at least one processor 604 that is connected to a bus 602, wherein the computer system 600 may be implemented using any suitable protocol, such as PCI (Peripheral Component Interconnect), PCI-Express, AGP (Accelerated Graphics Port), Hyper Transport, or any other bus or point-to-point communication protocol (s). The computer system 600 also includes a memory 606.

[0110] Control logic (software) and data are stored in the memory 606 which may take a form of random-access memory (RAM). In the disclosure, a single semiconductor platform may refer to a sole unitary semiconductor-based integrated circuit or chip. It should be noted that the term single semiconductor platform may also refer to multi-chip modules with increased connectivity which simulate on-chip operation, and make substantial improvements over utilizing a conventional central processing unit (CPU) and bus implementation. Of course, the various modules may also be situated separately or in various combinations of semiconductor platforms per the desires of the user. The computer system 600 may also include a secondary storage 610. The secondary storage 610 includes, for example, a hard disk drive and a removable storage drive, representing a floppy disk drive, a magnetic tape drive, a compact disk drive, digital versatile disk (DVD) drive, recording device, universal serial bus (USB) flash memory. The removable storage drive at least one of reads from and writes to a removable storage unit in a well-known manner.

[0111] Computer programs, or computer control logic algorithms, may be stored in at least one of the memory 606 and the secondary storage 610. Such computer programs, when executed, enable the computer system 600 to perform various functions as described in the foregoing. The memory 606, the secondary storage 610, and any other storage are possible examples of computer-readable media.

[0112] In an implementation, the architectures and functionalities depicted in the various previous figures may be implemented in the context of the processor 604, a graphics processor coupled to a communication interface 612, an integrated circuit (not shown) that is capable of at least a portion of the capabilities of both the processor 604 and a graphics processor, a chipset (namely, a group of integrated circuits designed to work and sold as a unit for performing related functions, and so forth).

[0113] Furthermore, the architectures and functionalities depicted in the various previous-described figures may be implemented in a context of a general computer system, a circuit board system, a game console system dedicated for entertainment purposes, an application-specific system. For example, the computer system 600 may take the form of a desktop computer, a laptop computer, a server, a workstation, a game console, an embedded system.

[0114] Furthermore, the computer system 600 may take the form of various other devices including, but not limited to a personal digital assistant (PDA) device, a mobile phone device, a smart phone, a television, and so forth. Additionally, although not shown, the computer system 600 may be coupled to a network (for example, a telecommunications network, a local area network (LAN), a wireless network, a wide area network (WAN) such as the Internet, a peer-to-peer network, a cable network, or the like) for communication purposes through an I / O interface 608

[0115] It should be understood that the arrangement of components illustrated in the figures described are exemplary and that other arrangement may be possible. It should also be understood that the various system components (and means) defined by the claims, described below, and illustrated in the various block diagrams represent components in some systems configured according to the subject matter disclosed herein. For example, one or more of these system components (and means) may be realized, in whole or in part, by at least some of the components illustrated in the arrangements illustrated in the described figures. In addition, while at least one of these components are implemented at least partially as an electronic hardware component, and therefore constitutes a machine, the other components may be implemented in software that when included in an execution environment constitutes a machine, hardware, or a combination of software and hardware.

[0116] Although the disclosure and its advantages have been described in detail, it should be understood that various changes, substitutions, and alterations can be made herein without departing from the spirit and scope of the disclosure as defined by the appended claims.

Claims

CLAIMS1. A method of correcting junction shapes in a road network map, the method comprising: obtaining crosswalk features by a classifier (212) from an aerial image (208, 308) of a road and an authoritative road network map (210, 310) of the road, creating an enhanced road network map (320) by co-locating the crosswalk features as crosswalk instances (312) into the authoritative road network map, wherein each crosswalk instance in the enhanced road network map comprises coordinates of a boundary (316) of a single crosswalk, grouping the crosswalk instances of the enhanced road network map into one or more junction instances by a geometrical reasoner, refining the coordinates of the boundaries of the crosswalk instances grouped into the junction instances based on coordinate centers of the junctions from the authoritative road network map by the geometrical reasoner, updating road network nodes of the enhanced road network map by repositioning and / or removing one or more road network nodes of the authoritative road network map and / or adding one or more new road network nodes to the enhanced road network map by the geometrical reasoner, updating road network segments of the enhanced road network map by one or more of repositioning, resizing and removing one or more road network segments of the authoritative road network map in accordance with the updated road network nodes by the geometrical reasoner, and updating attributes of the updated road network nodes and the updated road network segments in the enhanced road network map to provide for a connectivity of the enhanced road network by the geometrical reasoner.

2. The method of claim 1, further comprising: converting the enhanced road network map into a pre-defined road network format (230) for a road traffic simulator.

3. The method of claim 1 or 2, wherein the refining of the coordinates of the boundaries of the crosswalk instances grouped into the junction instances comprises shifting the boundariesof the crosswalk instances grouped into each junction instance with respect to a coordinate center of the junction from the authoritative road network map.

4. The method of any of claims 1 to 3, wherein the obtaining of crosswalk features from the aerial image of the road comprises retrieving image features relevant for crosswalks by the classifier with a Deep Learning, DL, segmentation algorithm.

5. The method of claim 4, further comprising: training a DL model of the DL segmentation algorithm based on input aerial images of roads with crosswalks, centered on positions of road network nodes associated with a junction of the authoritative road network map and annotated pixel-wise into classes of crosswalks and background components, using transfer learning starting from a pre-trained model with applying different types of data augmentation to the input images.

6. The method of any of claims 1 to 3, wherein the obtaining of crosswalk features from the aerial image of the road comprises retrieving image features by the classifier with a semantic segmentation method and instantiating single crosswalk objects.

7. The method of any of claims 1 to 6, further comprising: providing map features for the enhanced road network map by a user, wherein the map features include one or more of road network instances of different types, attributes, and natural terrain characteristics.

8. The method of any of claims 1 to 7, wherein the creating of the enhanced road network map comprises co-locating the crosswalk features as crosswalk instances into the authoritative road network map using coordinate projection and semantic assignment.

9. The method of any of claims 1 to 8, wherein the grouping of the crosswalk instances of the enhanced road network map into the junction instances is based on one or more of semantic information about a road network node on which the input aerial image was centered, a proximity of a crosswalk to a nearest junction, and a junction importance defined by a number of zebra crossings associated with the junction.

10. The method of any of claims 1 to 9, wherein the refining of the coordinates of the boundaries of the crosswalk instances grouped into each junction instance comprises:shifting the coordinates of the boundaries of the grouped crosswalk instances all together to move their coordinate center into a coordinate center of the junction according to the authoritative road network map, and shifting the coordinates of the boundaries of one or more of the grouped crosswalk instances horizontally and / or vertically to center their positions with respect to positions of the road network segments in the enhanced road network map.

11. The method of any of claims 1 to 8, wherein the grouping of the crosswalk instances of the enhanced road network map into the junction instances and / or the refining of the coordinates of the boundaries of the crosswalk instances grouped into each junction instance comprises using one or more of a machine learning, ML-based model, a distance-based heuristic model, a Neural Network, and a non-maximal suppression algorithm.

12. The method of any of claims 1 to 11, wherein the updating of road network nodes of the enhanced road network map comprises, for each junction instance: generating a junction boundary to enclose the boundaries of all the crosswalk instances grouped into the junction instance, determining cut points between edges of the junction boundary and the boundaries of the crosswalk instances grouped into the junction instance using a spatial index, and one or more of repositioning all road network nodes of the enhanced road network map located within the junction boundary to edges of the junction boundary, removing one or more road network nodes within the junction boundary, and adding one or more new road network nodes outside the junction boundary using the cut points.

13. The method of any of claims 1 to 12, wherein the updating of road network nodes of the enhanced road network map further comprises determining updated positions of the road network nodes based on an information about one or more of lane numbers for different moving directions of the road, a road and / or lane width, and internal junction markings.

14. The method of claim 12 or 13, wherein the updating of road network segments of the enhanced road network map comprises, for each junction instance:repositioning a starting or an ending point of each road network segment related to the junction instance and connected to one of the road network nodes originally located within the junction boundary, so that the repositioned starting or ending point coincides with an updated position of said road network node at the edges of the junction boundary, wherein parts of the repositioned road network segments originally located within the junction boundary are removed, and translating each road network segment that connects two road network nodes originally located within the junction boundary to a segment outside of the junction boundary, so that predecessors and successors of said two road network nodes are preserved and a length of the translated road network segment is resized to be proportional to an original length of the road network segment within the junction boundary.

15. The method of claim 14, wherein the updating of attributes of the road network nodes and the road network segments in the enhanced road network map comprises: updating metadata of the enhanced road network map in accordance with updates of the road network nodes, and updating metadata of the enhanced road network map in accordance with updates of the road network segments.

16. The method of any of claims 1 to 15, further comprising: correcting lane connections in the enhanced road network map with the updated road network nodes and the updated road network segments by one or more of identifying successor and predecessor lanes for a given road network segment, removing overlapping nodes and determining permitted movement directions at intersections.

17. An apparatus for correcting junction shapes in a road network map, the apparatus comprising: an input module (102, 202, 302) configured for: obtaining crosswalk features with a classifier (212) from an aerial image (208,creating an enhanced road network map (320) by co-locating the crosswalk features as crosswalk instances (312) into the authoritative road network map, wherein each crosswalk instance in the enhanced road network map comprises coordinates of a boundary (316) of a single crosswalk, a geometrical reasoner processing module (104, 204, 304) configured for: grouping the crosswalk instances of the enhanced road network map into one or more junction instances, refining the coordinates of the boundaries of the crosswalk instances grouped into the junction instances based on coordinate centers of the junctions from the authoritative road network map, updating road network nodes of enhanced road network map by repositioning and / or removing one or more road network nodes of the authoritative road network map and / or adding one or more new road network nodes to the enhanced road network map, updating road network segments of the enhanced road network map by one or more of repositioning, resizing and removing one or more road network segments of the authoritative road network map in accordance with the updated road network nodes, and updating attributes of the updated road network nodes and the updated road network segments in the enhanced road network map to provide for a connectivity of the enhanced road network.

18. The apparatus of claim 17, further comprising an output module (106, 206, 306) configured for converting the enhanced road network map into a pre-defined road network format (230) for a road traffic simulator.

19. The apparatus of claim 17 or 18, wherein the geometrical reasoner processing module is configured for shifting the boundaries of the crosswalk instances grouped into each junction instance with respect to a coordinate center of the junction from the authoritative road network map.

20. The apparatus of any of claims 17 to 19, wherein the classifier is configured for retrieving image features relevant for crosswalks from the aerial image of the road with a Deep Learning, DL, segmentation algorithm.

21. The apparatus of claim 20, wherein the input module is configured for training a DL model of the DL instance segmentation algorithm based on input aerial images of roads with crosswalks, centered on positions of road network nodes associated with a junction of the authoritative road network map and annotated pixel-wise into classes of crosswalks and background components, using transfer learning starting from a pre-trained model with applying different types of data augmentation to the input images.

22. The apparatus of any of claims 17 to 19, wherein the classifier is configured for retrieving image features from the aerial image of the road with a semantic segmentation method and for instantiating single crosswalk objects.

23. The apparatus of any of claims 17 to 22, wherein the input module is further configured for obtaining map features for the enhanced road network map provided by a user, wherein the map features include one or more of road network instances of different types, attributes, and natural terrain characteristics.

24. The apparatus of any of claims 17 to 23, wherein the input module is configured for colocating the crosswalk features as crosswalk instances into the authoritative road network map using coordinate projection and semantic assignment.

25. The apparatus of any of claims 17 to 24, wherein the geometrical reasoner processing module is configured for grouping of the crosswalk instances of the enhanced road network map into the junction instances based on one or more of semantic information about a road network node on which the input aerial image was centered, a proximity of a crosswalk to a nearest junction, and a junction importance defined by a number of zebra crossings associated with the junction.

26. The apparatus of any of claims 17 to 25, wherein the geometrical reasoner processing module is configured for refining the coordinates of the boundaries of the crosswalk instances grouped into each junction instance by means ofshifting the coordinates of the boundaries of the grouped crosswalk instances all together to move their coordinate center into a coordinate center of the junction according to the authoritative road network map, and shifting the coordinates of the boundaries of one or more of the grouped crosswalk instances horizontally and / or vertically to center their positions with respect to positions of the road network segments in the enhanced road network map.

27. The apparatus of any of claims 17 to 24, wherein the geometrical reasoner processing module is configured for grouping the crosswalk instances of the enhanced road network map into the junction instances and / or refining the coordinates of the boundaries of the crosswalk instances grouped into each junction instance by means of using one or more of a machine learning, ML-based model, a distance-based heuristic model, a Neural Network, and a non- maximal suppression algorithm.

28. The apparatus of any of claims 17 to 27, wherein the geometrical reasoner processing module is configured for updating road network nodes of the enhanced road network map, for each junction instance, by means of generating a junction boundary to enclose the boundaries of all the crosswalk instances grouped into the junction instance, determining cut points between edges of the junction boundary and the boundaries of the crosswalk instances grouped into the junction instance using a spatial index, and one or more of repositioning all road network nodes of the enhanced road network map located within the junction boundary to edges of the junction boundary, removing one or more road network nodes within the junction boundary, and adding one or more new road network nodes outside the junction boundary using the cut points.

29. The apparatus of any of claims 17 to 28, wherein the geometrical reasoner processing module is further configured for determining updated positions of the road network nodes based on an information about one or more of lane numbers for different moving directions of the road, a road and / or lane width, and internal junction markings.

30. The apparatus of claim 28 or 29, wherein the geometrical reasoner processing module is configured for updating road network segments of the enhanced road network map, for each junction instance, by means of repositioning a starting or an ending point of each road network segment related to the junction instance and connected to one of the road network nodes originally located within the junction boundary, so that the repositioned starting or ending point coincides with an updated position of said road network node at the edges of the junction boundary, wherein parts of the repositioned road network segments originally located within the junction boundary are removed, and translating each road network segment that connects two road network nodes originally located within the junction boundary to a segment outside of the junction boundary, so that predecessors and successors of said two road network nodes are preserved and a length of the translated road network segment is resized to be proportional to an original length of the road network segment within the junction boundary.

31. The apparatus of claim 30, wherein the geometrical reasoner processing module is configured for updating attributes of the road network nodes and the road network segments by means of updating metadata of the enhanced road network map in accordance with updates of the road network nodes, and updating metadata of the enhanced road network map in accordance with updates of the road network segments.

32. The apparatus of any of claims 17 to 31, wherein the geometrical reasoner processing module is further configured for correcting lane connections in the enhanced road network map with the updated road network nodes and the updated road network segments by one or more of identifying successor and predecessor lanes for a given road network segment, removing overlapping nodes and determining permitted movement directions at intersections.

Images