Thicket data repository for the EEG
anil.recoil.org
1{
2 "id": "https://gabrielmahler.org/walkability/compsci/2025/06/03/related%20work",
3 "title": "Walkability Chapter 3: Related Work",
4 "link": "https://gabrielmahler.org/walkability/compsci/2025/06/03/related-work.html",
5 "updated": "2025-06-03T10:40:11",
6 "published": "2025-06-03T10:40:11",
7 "summary": "Related work",
8 "content": "<h1>Related work</h1>\n\n<p>To profile existing work relevant to urban pedestrian path-finding, we analyze three important areas of research and development -\nstate-of-the-art pedestrian routing approaches, available\nsources of geospatial data, and modern data-centered\napproaches for analysis and querying of urban areas.\nFurthermore, we highlight the lack of overlap between the pedestrian\nrouting domain and the domain of geospatial assessments.</p>\n\n<h2>Transportation routing</h2>\n\n<p>Path‑finding has been explored in great depth within the field of human\ntransportation, where route‑planning must address an ever‑expanding list\nof considerations, from static geospatial information to constantly\nchanging real‑time inputs like traffic flows or weather. These\nchallenges are often rooted in the diversity of path-finding\nrequirements and objectives introduced by the spectrum of transportation\nmodes. Different modes introduce different infrastructure network\nmodifications, performance requirements, and external factors\nconsiderations.</p>\n\n<p>Nonetheless, as discussed in the following paragraphs, despite\nimplementing these complex considerations, path-finding networks\ngenerally allow for little or no input related to factors other than\nbasic information about the transportation segments (such as the\nsegments’ surfaces) or elementary geospatial raster data (such as\nelevation profiles). Furthermore, the focus overwhelmingly lies within\nthe maximization of temporal efficiencies of the generated paths. If\nsomeone were to use these frameworks for the generation of personalized\npaths (such as ones more relatable to the concepts of walkability), they\nwould have to leverage these primitive parameters based on their\nindirect correlation with the desired characteristics.</p>\n\n<h3>Open-Source Path-Finding</h3>\n\n<p>Open‑source initiatives have played a significant role in the\npath-finding domain, and pushed its development forward, often by\nleveraging the vast amounts of freely available geospatial datasets,\nsuch as those provided by the publicly-maintained OpenStreetMap project. Popular\nframeworks such as Valhalla (Valhalla Contributors 2025), Open Source\nRouting Machine (OSRM) (Luxen and Vetter 2011), and\nGraphHopper (Contributors 2025) employ a mixture of path-finding\nalgorithms and address the diversity of possible requirements (such as\nthe different modes of transportation) through customizable\nconfigurations. Such configurations frequently focus on, for instance,\ndefining assumptions about travel speeds on specific physical surfaces.\nThe frameworks then use these definitions to readjust the transportation\nnetworks’ cost distributions.</p>\n\n<p>Both GraphHopper (Contributors 2025) and OSRM (Luxen and Vetter 2011)\nexecute the network tuning with pre-defined “profiles”. After the\nadjustment, they rely on a mix of multi‑level search algorithms, graph\ncontraction techniques, and bidirectional heuristic functions. In\nGraphHopper, profiles can configure mappings to specific geospatial\nelements and construct mode-aware cost functions and objective\n“weightings”. For instance, a weighting can be such that it minimizes\nthe travel time or the travel distance. With these, GraphHopper can also\nbe configured to consider more specific factors, such as elevation or\ncurvature. These parameters are reflected both in the estimation of\nedges’ importance during the contraction hierarchies pre-processing and\nduring the search time. For the search, and depending on the specific\ncase, GraphHopper uses either uni-directional or bi-directional A* and\nDijkstra (Contributors 2025).</p>\n\n<p>An approach similar to GraphHopper is implemented by OSRM (Luxen and\nVetter 2011). Like GraphHopper, OSRM uses predefined “profiles”, which\ndefine the anticipated speeds and eligibility criteria for specific\nsegment types. Additionally, OSRM also supports setting various penalty\nflags for specific features within the network, such as intersections or\nU-turns. OSRM by default relies on a multi-level variation of the\nDijkstra algorithm (derived from <em>Delling et. al.</em> (Delling et al. 2017;\nLuxen and Vetter 2011)). Using the pre-calculated matrices, OSRM then\nbuilds a multi-level partitioned graph and initiates the search. It is\ngenerally only in instances where the problem’s distance matrices grow\nvery large that contraction hierarchies are used (Luxen and Vetter\n2011).</p>\n\n<p>Valhalla presents a slightly different approach by using dynamic costing\nfunctions, which are applied during the search’s runtime. Rather than\nrelying on precomputed weights in a graph, Valhalla computes the costs\nof traversing through segments and junctions dynamically, while leaving\nthe underlying segment network unmodified. The specifications of\ndifferent mode profiles are, therefore, defined by the costing method\nand its parameters. In Valhalla, the logic of these costing methods is\nimplemented in a similar fashion to OSRM or GraphHopper. Elements such\nas elevation, traffic rules, junctions, or segments’ surfaces are\nleveraged to define user preferences. Then, during the search phase, the\nlogic defined by the selected costing class is injected into the usual\nroutine of the unidirectional or bidirectional A* (Valhalla\nContributors 2025).</p>\n\n<p>All three of these open-source frameworks allow for some degree of\npreference input, either through modifying the static graphs of segments\nor by parameterizing the cost calculations during runtime. These\nframeworks are (to some limited degree) capable of reflecting sets of\nspecific preferences in their route generation, by filtering outright\nineligible options (such as highways for walking), and by rewarding or\npenalizing specific features of segments and junctions (such as their\nsurfaces). Nevertheless, they require manually constructed definitions\nof the preferences, and are generally built upon only a fraction of the\navailable data, ignoring much of the full geospatial context.</p>\n\n<p>In contrast to these open-source frameworks, <em>Socharoentum et. al.</em>\nattempts to integrate more context into pedestrian path\ngeneration (Socharoentum and Karimi 2016). The underlying motivation\nlies in acknowledging that pedestrian route-finding is more than a\nsimple shortest-path problem, but requires a careful consideration of\nother contextual factors as well. Therefore, they decompose each edge of\nthe pedestrian network into data points that reflect a larger set of\nattributes (such as slope variation, air pollution level, and estimated\nshade levels) so that each pedestrian segment can be evaluated against\nmultiple criteria and recombined into an exact composite score for route\nselection. Unlike the afore-discussed frameworks, they incorporate more\nthan only the factors strictly associated with the infrastructural\nsegments themselves, and have, in experimental results, shown an ability\nto reveal more pedestrian-friendly walking paths, rather than only the\nquickest ones. Nevertheless, it is still very limited by its dependence\non static settings (which define the user preferences and the\nalgorithm’s parameters) and a narrow selection of the considered\nfactors.</p>\n\n<h3>Proprietary Route-Finding Algorithms</h3>\n\n<p>In terms of sheer popularity, open-source path-finding frameworks can\nhardly compare with the proprietary alternatives, such as Google\nMaps (Schröder-Bergen 2024). There is, however, generally little known\nabout the inner workings of these services, and theories about the\nspecific implementations and routing strategies are mostly based on\nspeculations. Google Maps is believed to rely on a familiar mixture of\ncontraction hierarchies, and Dijkstra and A* variants (Deshmukh 2025).\nUnlike the open-source services, however, Google Maps leverages immense\namounts of data (including data from its proprietary crowd-sourcing\nstreams) in graph-based machine learning models. These models are\nleveraged in conjunction with other methods to not only generate paths\nbased on existing conditions (for example, historical patterns,\ngeospatial information, or user feedback) but also to predict future\nevents, such as possible traffic jams. Therefore, Google Maps utilizes a\ntype of dynamic costing. Moreover, it is apparent that Google Maps uses\nsome equivalent of the open-source routing profiles. For instance,\nGoogle Maps can optionally provide “eco-friendly” automotive routes,\nestimated based on types of roads, elevation profiles, or junction types\nand frequencies (Moondra, n.d.). Nevertheless, it is difficult to deduce\nmuch about the routing logic Google Maps implements. It is also unclear\nwhether or not Google Maps reserves some of its tools (for instance, its\ngraph-based machine learning algorithms) only for specific\ntransportation modes, such as automotive transportation (Lau 2020;\nDeshmukh 2025).</p>\n\n<h3>Machine Learning-centered Approaches</h3>\n\n<p>The recent popularity of machine learning-based solutions has also been\nreflected in the path-finding domain, and several research projects have\naimed to optimize or expand the capabilities of the traditional\npath-finding approaches.</p>\n\n<p><em>Wang et. al.</em> presented an approach which eliminates the need for\nmanually crafted heuristic and cost functions by replacing them with\ntrainable neural networks (J. Wang et al. 2019). They suggest using a\nRecurrent Neural Network (RNN) to model the expected cost from the\ninitial to the currently-iterated state. The proposed RNN calculates the\nlog-negative probability of a user taking a step, while considering a\nvectorized representation of the user, the current location, and the\ntemporal context. To enhance A*’s ability to provide accurate estimates\nof the remaining cost (which is often calculated with primitive methods\nsuch as Euclidean distance), they suggest leveraging a moving state of\nthe user in a value model, built upon a graph attention network, to\ncapture the structural characteristics. The validity of this approach\nwas demonstrated with several experiments, where raw GPS logs were\nmatched onto a segment graph built from OpenStreetMap data, and each trip\nwas converted into a sequence of junction visits. Vectorized\nrepresentations combining unique embedding vectors and discretized\ntemporal features (specifically, the day of the week and the hour of the\nday) were then assigned to each user, and at each junction, concatenated\nwith a location embedding. These embeddings then served as an input into\nthe RNN, and the moving‐state output fed into the value network.</p>\n\n<p>For some time, there has also existed an overlap between path-finding\nand natural language processing. Due to their integral predispositions,\nLLMs do not offer only an effective way to parse and process large\nvolumes of various textual data, but also provide a degree of\n“reasoning”. This reasoning can be particularly useful for evaluating\nvarious contexts against some requirements specified by, for instance,\nuser queries.</p>\n\n<p>A relatively straightforward way to leverage the potential of LLMs is to\nintegrate them into retrieval augmented generation (RAG) systems. These\nsystems improve the generation capabilities of LLMs by extracting\nrelevant data from external knowledge bases and providing it as context\nto the LLMs, thus extending their knowledge and minimizing\nhallucinations (Gao et al. 2023).</p>\n\n<p><em>Marcelyn et. al.</em>, for instance, introduce a RAG-like approach adjusted\nspecifically for path-generation (Marcelyn et al. 2025). They aim to\nleverage the versatility of LLMs’ reasoning to produce personalized\nroutes without any further training. During the initialization phase,\nthe framework queries geospatial datasets and processes sets of\ntrajectories of historical paths into concise pieces of human-readable\ndescriptions of the shortest, fastest, and most frequently taken routes\nbetween two destinations. These descriptions are then encoded into\nembedding vectors and stored in a vector database. During inference, a\nuser query is embedded with the same encoder and used for comparison\nagainst the stored vectorized routes, the closest few of which are\nretrieved and injected into a prompt passed into an LLM. The LLM is then\nrun to consider the compiled context and returns a sequence of segments\nthat aim to maximize the similarity to the user query. This way, they\nallow for a certain degree of user preference input, but rely heavily on\nhistorical data to perform route planning, and only a restricted\nselection of the full geospatial context.</p>\n\n<h2>Automated Walkability Assessments</h2>\n\n<p>Similarly to pedestrian path-finding frameworks, tendencies to omit the\nfull geospatial context are also evidenced by the limited volume of\nresearch focused on exploring automated evaluation of walkability\nthrough computer science methods. Although it is apparent that existing\nservices and tools used for automated walkability assessments (such as\nthe NWI or WalkScore) are, overall, insufficient, there hasn’t been much\nacademic research into exploring ways to replace them. Besides a few\nproposed methods (for example, published by <em>Horak et. al.</em> (Horak et\nal. 2022)) that leverage the concept of “<em>n</em>-minute cities”, few projects have\nexplored assessing infrastructure and physical environments through\nmodern computing and statistical methods (such as machine learning).</p>\n\n<p>The main direction in the few existing computer science-centered\nprojects has been leveraging vision networks, particularly on imaging\ndata aggregated from street-view datasets. A number of published\napproaches employ a general workflow that systematically samples and\nprocesses street-view imagery through machine learning-based\nsegmentation to derive quantitative indicators relevant to evaluating\nthe environment from a pedestrian perspective (Yin and Wang 2016;\nMiranda et al. 2021; Z. Huang et al. 2024; X. Huang et al. 2024).\nBroadly, these methods collect images in regular intervals along street\nnetworks and then conduct a pixel-level analysis via various\nsegmentation neural networks to delineate relevant environmental\nelements. The extracted features are subsequently aggregated into\npredefined indices and integrated with complementary spatial or\nbehavioral analyses (e.g., space syntax integration values, deviations\nfrom shortest-path trajectories in GPS data, pedestrian counts, or\nfacility accessibility) to validate various properties, including their\nrelevance to determining walkability.</p>\n\n<p>Reflecting the advancements of image-based neural networks over the past\nseveral years, these visual models are very capable of assessing the\nphysical environment and providing insights valuable for conducting\nwalkability assessments. Nonetheless, street-view imagery generally\nsuffers from limited temporal representativeness, as it is infrequently\nand inconsistently updated. Imaging datasets necessary for covering\nentire cities and regions are also demanding to upkeep and aggregate.\nThese frameworks may also often overemphasize strictly visual features.\nFinally, using composite indices can be susceptible to generic scoring,\nassuming universal preferences that are incompatible with specific\nregions.</p>\n\n<h2>Geospatial Datasets</h2>\n\n<p>Imaging datasets, fortunately, do not represent the only modality\nrelevant to geospatial tasks. In fact, in recent years, large volumes of\ntext-based geospatial datasets have been made available to the public\nboth through publicly and privately managed initiatives. Furthermore,\nseveral of these data sources are spatially universal and aim to record\ndata from the entire world.</p>\n\n<h3>Centrally-managed Datasets</h3>\n\n<p>Perhaps unexpectedly, several such data sources have been enabled by\nmajor private data collectors. Overture Maps (Overture Maps Foundation\n2025) is one such example, and has received much attention for its\npublic mission (Sawers 2024). It was founded to oppose Google Maps’\ndominance in geographic data aggregation, and has since relied on data\ncontributed by its core members and external contributors, such as Meta,\nMicrosoft, AWS, or Tripadvisor. Furthermore, Overture Maps has also\nextracted data from governmental and open datasets. The aggregated\ndataset contains information about several feature types, including\nbuildings, geographical boundaries, and various points of interest.</p>\n\n<p>In contrast, the Foursquare dataset (Foursquare Labs, Inc. 2025) is\nfocused exclusively on commercial places of interest, such as\nestablishments or offices. Furthermore, Foursquare aggregates its data\nfrom a diverse variety of sources (Foursquare 2021). Large amounts of\ndata come from open web crawling, during which Foursquare scrapes\ndirectory websites or chain locator pages. Some data is also extracted\nfrom third-party licensors, other data representatives, and datasets\ncurated by various communities or outsourced teams.</p>\n\n<p>Nonetheless, based on a manual review conducted as part of this work\n(limited to the boundaries of Cambridge, UK), both Overture Maps and\nFoursquare exhibit vast temporal inaccuracy and undesirable data\nfiltering. Both datasets include numerous entries that no longer exist\n(such as closed establishments or temporary food tents), and omit points\nimportant for estimating walkability (such as public amenities or\naesthetic-inducing and accessibility elements).</p>\n\n<h3>Crowd-sourced Datasets</h3>\n\n<p>However, in addition to centrally managed datasets, the geospatial data\ndomain has also long involved crowd-sourcing projects and initiatives,\nwhich, across many regions of the world, have resulted in vast and\ndiverse data sources. OpenStreetMap (OSM) (OpenStreetMap contributors\n2017) is a central pillar of this movement.</p>\n\n<p>OSM is a free, open‐content map dataset of the world, and is maintained\nby a global community of volunteer mappers, who gather geographic\ninformation through various field surveys (often relying on GPS\ntracking), aerial and satellite imagery tracing, and aggregation of\nother freely licensed sources. Contributions are collected into a\ncentralized database and released under the loosely restrictive Open\nDatabase License.</p>\n\n<p>OSM achieves a high versatility of the geospatial data it can support by\nrelying on only a few simple data types. “Nodes” (for singular points),\n“ways” (for ordered sequences of nods), and “relations” (for groupings\nof nodes and ways and constructing more complex map features) are the\nthree geometric primitives OSM leverages, and which, extended with sets\nof textual key-value pairs, can supply extremely high descriptiveness\nand spatial accuracy. Hence, OSM regularly contains information about\nall sorts of various urban elements (such as artworks, various\ninfrastructure, or street furniture), which other datasets do not\ninclude or could not even support. Another unique strength of OSM is its\nconnectivity with other online databases, such as Wikidata - a large\npublic knowledge base that also includes entries of places, buildings,\nand streets.</p>\n\n<p>Nevertheless, while in many parts of the world OSM provides a highly\ndescriptive and temporally accurate coverage supported by active\ncommunities of contributors, in some regions, OSM’s coverage is far from\nsatisfying. <em>Herfort et. al.</em> conclude in a recent analysis that in a\nconsiderable sample of global cities, OSM’s coverage of building\nfootprints exceeded 80% completeness in 14% of the analyzed cities, but\nfalls short of 20% completeness in 69% of the sample (Herfort et al.\n2023). This inconsistency across cities is related to the respective\ngeographical context, as the completeness remains relatively high in\nEurope and Central Asia (71%), and North America (64%), but is\nconsiderably lower in Latin America & the Caribbean (20%), East Asia &\nPacific (20%), Middle East and North Africa (12%), and South Asia (9%).\nOSM’s daily contribution statistics (monitored in the online OSMstats\nproject (Neis 2025)) only underline these trends. In the long term,\nEuropean and North American countries (such as Germany, the United\nStates, France, and the United Kingdom) are associated with\nsignificantly larger cohorts of individual contributors and receive\ndisproportionately higher numbers of contributions.</p>\n\n<p>Nonetheless, disproportionately better coverage of Western countries is\na global problem in the mapping domain, equally concerning corporate\ndatasets (and their dependents - such as Overture Maps (Overture Maps\nFoundation 2025)) (Tockner et al. 2025).</p>\n\n<h3>Local Administration Datasets</h3>\n\n<p>Lastly, some governmental offices and national agencies aggregate and\nprovide public access to various geospatial datasets. These datasets\ntend to be more specialized and frequently tied to specific regions. To\naddress the importance of aesthetics and nature underlined by the\nwalkability-related urbanist literature, two geospatial\nfeature types were particularly relevant in the context of our work:\ngreenery elements and public green spaces.</p>\n\n<p>While the level of greenery in a physical environment can be estimated\nbased on a variety of factors, trees and large shrubs were assumed to be\nof particular importance, partly because of the climate and urbanist\ncharacter in our experimental region of interest - cities in the UK. To\ngather information about trees, moreover, several official\nagency-provided sources were available. For instance, the Cambridge City\nCouncil regularly releases a carefully curated collection of trees in\nCambridge (Council 2024). However, a more comprehensive picture of the\ntree cover can be created based on a LIDAR Vegetation Object Model (VOM)\npublished by the Department of Food, Environment & Rural Affairs Data\nServices Forum (Agency 2023). VOM is a raster dataset aimed at\nidentifying existing riparian tree cover and future opportunities for\ntree planting. It was generated through modeling of the Environment\nAgency’s 1m elevation LIDAR data, with the classification of vegetation\nand non-vegetation objects supported by spatial proximity analysis using\nvarious mappings and Sentinel-2 imagery.</p>\n\n<p>Another relevant source is the Open Greenspace dataset, released by the\nOrdinance Survey (UK’s national mapping agency) (Survey 2025). The aim\nof Open Greenspace is to improve the accessibility and awareness of\nareas, such as parks and sports facilities, which are likely available\nto the public. Open Greenspace is updated regularly by the Ordinance\nSurvey every six months.</p>\n\n<h2>Data Integration in Geographic Information Systems</h2>\n\n<p>The growing abundance of publicly accessible geospatial data has been\nleveraged by various data-centered frameworks, particularly in the realm\nof natural language processing.</p>\n\n<p>Of particular interest has been the augmentation of LLMs with\ncartographic and spatial data to create interactive mapping systems and\nspatial intelligence frameworks. Various research projects focused on\nsteering LLMs toward spatial reasoning by translating geospatial data\n(represented in different mixtures of map features, points of interest,\nroad networks, and street‐level imagery) into text or low-dimensional\nembeddings (Unlu 2023; Manvi et al. 2023; Balsebre et al. 2023;\nBalsebre, Huang, and Cong 2024; Li et al. 2024). Generally, these works\nhave applied some form of fine-tuning or contrastive multi-modal\ntraining to align their models’ internal knowledge with the geospatial\ndata. The aim has been to enable the LLMs to learn to answer questions\nabout urban areas, plan itineraries, or infer things like traffic speeds\nor population densities. By layering various models and tools, these\narchitectures have been able to draw on both visual and textual cues to\nextract and reason about geographic information without having to\nconstruct entirely new spatial systems.</p>\n\n<p>Nonetheless, despite its potential, the use of vectorized geospatial\nrepresentations for other purposes has been sparsely explored. One of\nthe few exceptions is presented by <em>Tempelmeier et. al.</em>, who build a\ncomprehensive dataset of geospatial entity embeddings to minimize the\nchallenges of using OSM data in machine learning applications due to its\nvast scale, inconsistent tagging, and lack of structured\nsemantics (Tempelmeier, Gottschalk, and Demidova 2021). They address\nthese issues by generating embeddings of OSM entities with their own\nencoder, capturing spatial location and semantic representation of the\navailable data. This approach’s usefulness is later demonstrated in an\nassertion task (where the embeddings were leveraged to assign Wikidata\nclasses to the OSM entities), and a link prediction task (where\ncountries of origin of OSM entities were predicted).</p>\n\n<h3>References</h3>\n\n<ul>\n <li>Environment Agency. (2023). <em>Lidar Vegetation Object Model (VOM)</em>. <em>Defra Data Services Platform</em>. Department for Environment, Food & Rural Affairs.\n<a href=\"https://environment.data.gov.uk/dataset/ecae3bef-1e1d-4051-887b-9dc613c928ec\">https://environment.data.gov.uk/dataset/ecae3bef-1e1d-4051-887b-9dc613c928ec</a></li>\n <li>Balsebre, P., Huang, W., & Cong, G. (2024). <em>LAMP: A Language Model on the Map</em>. <em>arXiv Preprint arXiv:2403.09059</em>.</li>\n <li>Balsebre, P., Huang, W., Cong, G., & Li, Y. (2023). <em>Cityfm: City Foundation Models to Solve Urban Challenges</em>. <em>arXiv Preprint arXiv:2310.00583</em>.</li>\n <li>GraphHopper Contributors. (2025). <em>GraphHopper Routing Engine</em>.\n<a href=\"https://github.com/graphhopper/graphhopper\">https://github.com/graphhopper/graphhopper</a></li>\n <li>Cambridge City Council. (2024). <em>Tree Data</em>.\n<a href=\"https://www.cambridge.gov.uk/tree-data\">https://www.cambridge.gov.uk/tree-data</a></li>\n <li>Delling, D., Goldberg, A. V., Pajor, T., & Werneck, R. F. (2017). <em>Customizable Route Planning in Road Networks</em>. <em>Transportation Science</em>, 51(2), 566–591.</li>\n <li>Deshmukh, A. (2025). <em>From A to B: Algorithms That Power Google Maps Navigation</em>. <em>Towards AI</em>.\n<a href=\"https://towardsai.net/p/artificial-intelligence/from-a-to-b-algorithms-that-power-google-maps-navigation\">https://towardsai.net/p/artificial-intelligence/from-a-to-b-algorithms-that-power-google-maps-navigation</a></li>\n <li>Foursquare. (2021). <em>How to Get POI Data Right</em>. <em>Foursquare Blog</em>.\n<a href=\"https://location.foursquare.com/resources/blog/leadership/how-to-get-poi-data-right\">https://location.foursquare.com/resources/blog/leadership/how-to-get-poi-data-right</a></li>\n <li>Foursquare Labs, Inc. (2025). <em>Foursquare Open Source Places Dataset</em>.\n<a href=\"https://opensource.foursquare.com/os-places/\">https://opensource.foursquare.com/os-places/</a></li>\n <li>Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., & Wang, H. (2023). <em>Retrieval-Augmented Generation for Large Language Models: A Survey</em>. <em>arXiv Preprint arXiv:2312.10997</em>, 2, 1.</li>\n <li>Herfort, B., Lautenbach, S., de Albuquerque, J. P., Anderson, J., & Zipf, A. (2023). <em>A Spatio-Temporal Analysis Investigating Completeness and Inequalities of Global Urban Building Data in OpenStreetMap</em>. <em>Nature Communications</em>, 14(1), 3985.</li>\n <li>Horak, J., Kukuliac, P., Maresova, P., Orlikova, L., & Kolodziej, O. (2022). <em>Spatial Pattern of the Walkability Index, Walk Score and Walk Score Modification for Elderly</em>. <em>ISPRS International Journal of Geo-Information</em>, 11(5), 279.</li>\n <li>Huang, X., Zeng, L., Liang, H., Li, D., Yang, X., & Zhang, B. (2024). <em>Comprehensive Walkability Assessment of Urban Pedestrian Environments Using Big Data and Deep Learning Techniques</em>. <em>Scientific Reports</em>, 14(1), 26993.</li>\n <li>Huang, Z., Wang, B., Luo, S., Wang, M., Miao, J., & Jia, Q. (2024). <em>Integrating Streetscape Images, Machine Learning, and Space Syntax to Enhance Walkability: A Case Study of Seongbuk District, Seoul</em>. <em>Land</em>, 13(10), 1591.</li>\n <li>Lau, J. (2020). <em>Google Maps 101: How AI Helps Predict Traffic and Determine Routes</em>. <em>Google</em>.\n<a href=\"https://blog.google/products/maps/google-maps-101-how-ai-helps-predict-traffic-and-determine-routes\">https://blog.google/products/maps/google-maps-101-how-ai-helps-predict-traffic-and-determine-routes</a></li>\n <li>Li, Z., Xu, J., Wang, S., Wu, Y., & Li, H. (2024). <em>StreetviewLLM: Extracting Geographic Information Using a Chain-of-Thought Multimodal Large Language Model</em>. <em>arXiv Preprint arXiv:2411.14476</em>.</li>\n <li>Luxen, D., & Vetter, C. (2011). <em>Real-Time Routing with OpenStreetMap Data</em>. In <em>GIS ‘11: Proceedings of the 19th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems</em> (pp. 513–516).\n<a href=\"https://doi.org/10.1145/2093973.2094062\">https://doi.org/10.1145/2093973.2094062</a></li>\n <li>Manvi, R., Khanna, S., Mai, G., Burke, M., Lobell, D., & Ermon, S. (2023). <em>GeoLLM: Extracting Geospatial Knowledge from Large Language Models</em>. <em>arXiv Preprint arXiv:2310.06213</em>.</li>\n <li>Marcelyn, S. C., Gao, Y., Zhang, Y., Gao, X., & Chen, G. (2025). <em>PathGPT: Leveraging Large Language Models for Personalized Route Generation</em>. <em>arXiv Preprint arXiv:2504.05846</em>.</li>\n <li>Miranda, A. S., Fan, Z., Duarte, F., & Ratti, C. (2021). <em>Desirable Streets: Using Deviations in Pedestrian Trajectories to Measure the Value of the Built Environment</em>. <em>Computers, Environment and Urban Systems</em>, 86, 101563.</li>\n <li>Moondra, M. (n.d.). <em>Navigate More Sustainably and Optimize for Fuel Savings with Eco-Friendly Routing – Google Maps Platform</em>. <em>Google</em>.\n<a href=\"https://mapsplatform.google.com/resources/blog/navigate-more-sustainably-and-optimize-fuel-savings-eco-friendly-routing/\">https://mapsplatform.google.com/resources/blog/navigate-more-sustainably-and-optimize-fuel-savings-eco-friendly-routing/</a></li>\n <li>Neis, P. (2025). <em>Statistics of the Free Wiki World Map</em>. <em>OSMstats</em>.\n<a href=\"https://osmstats.neis-one.org/\">https://osmstats.neis-one.org/</a></li>\n <li>OpenStreetMap Contributors. (2017). <em>Planet dump retrieved from <a href=\"https://planet.osm.org\">https://planet.osm.org</a></em>.\n<a href=\"https://www.openstreetmap.org\">https://www.openstreetmap.org</a></li>\n <li>Overture Maps Foundation. (2025). <em>Overture Maps Foundation</em>.\n<a href=\"https://overturemaps.org\">https://overturemaps.org</a></li>\n <li>Sawers, P. (2024). <em>Backed by Microsoft, AWS and Meta, the Overture Maps Foundation Launches Its First Open Map Datasets</em>. <em>TechCrunch</em>.\n<a href=\"https://techcrunch.com/2024/07/24/backed-by-microsoft-aws-and-meta-the-overture-maps-foundation-launches-first-open-map-datasets/\">https://techcrunch.com/2024/07/24/backed-by-microsoft-aws-and-meta-the-overture-maps-foundation-launches-first-open-map-datasets/</a></li>\n <li>Schröder-Bergen, S. (2024). <em>Google Maps and OpenStreetMap: The Coexistence of Two Unequal Siblings</em>. <em>Geoawesome</em>.\n<a href=\"https://geoawesome.com/google-maps-and-openstreetmap-the-coexistence-of-two-unequal-siblings\">https://geoawesome.com/google-maps-and-openstreetmap-the-coexistence-of-two-unequal-siblings</a></li>\n <li>Socharoentum, M., & Karimi, H. A. (2016). <em>Multi-Modal Transportation with Multi-Criteria Walking (MMT-MCW): Personalized Route Recommender</em>. <em>Computers, Environment and Urban Systems</em>, 55, 44–54.</li>\n <li>Ordinance Survey. (2025). <em>OS Open Greenspace</em>. <em>OS Download Products’ Documentation</em>.\n<a href=\"https://docs.os.uk/os-downloads/topography/os-open-greenspace\">https://docs.os.uk/os-downloads/topography/os-open-greenspace</a></li>\n <li>Tempelmeier, N., Gottschalk, S., & Demidova, E. (2021). <em>GeoVectors: A Linked Open Corpus of OpenStreetMap Embeddings on World Scale</em>. In <em>Proceedings of the 30th ACM International Conference on Information & Knowledge Management</em>, 4604–4612.</li>\n <li>Tockner, L., Herfort, B., Lautenbach, S., & Zipf, A. (2025). <em>Corporate Mapping in OpenStreetMap – Shifting Trends in Global Evolution and Small-Scale Effects</em>. <em>Geo-Spatial Information Science</em>, 1–20.</li>\n <li>Unlu, E. (2023). <em>Chatmap: Large Language Model Interaction with Cartographic Data</em>. <em>arXiv Preprint arXiv:2310.01429</em>.</li>\n <li>Valhalla Contributors. (2025). <em>Valhalla: Open Source Routing Engine for OpenStreetMap</em>.\n<a href=\"https://github.com/valhalla/valhalla\">https://github.com/valhalla/valhalla</a></li>\n <li>Wang, J., Wu, N., Zhao, W. X., Peng, F., & Lin, X. (2019). <em>Empowering A</em> Search Algorithms with Neural Networks for Personalized Route Recommendation*. In <em>Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining</em>, 539–547.</li>\n <li>Yin, L., & Wang, Z. (2016). <em>Measuring Visual Enclosure for Street Walkability: Using Machine Learning Algorithms and Google Street View Imagery</em>. <em>Applied Geography</em>, 76, 147–153.\n<a href=\"https://doi.org/10.1016/j.apgeog.2016.09.024\">https://doi.org/10.1016/j.apgeog.2016.09.02</a>.</li>\n</ul>",
9 "content_type": "html",
10 "author": {
11 "name": "",
12 "email": null,
13 "uri": null
14 },
15 "categories": [
16 "walkability",
17 "compsci"
18 ],
19 "source": "https://gabrielmahler.org/feed.xml"
20}