Related work

To profile existing work relevant to urban pedestrian path-finding, we analyze three important areas of research and development - state-of-the-art pedestrian routing approaches, available sources of geospatial data, and modern data-centered approaches for analysis and querying of urban areas. Furthermore, we highlight the lack of overlap between the pedestrian routing domain and the domain of geospatial assessments.

Transportation routing

Path‑finding has been explored in great depth within the field of human transportation, where route‑planning must address an ever‑expanding list of considerations, from static geospatial information to constantly changing real‑time inputs like traffic flows or weather. These challenges are often rooted in the diversity of path-finding requirements and objectives introduced by the spectrum of transportation modes. Different modes introduce different infrastructure network modifications, performance requirements, and external factors considerations.

Nonetheless, as discussed in the following paragraphs, despite implementing these complex considerations, path-finding networks generally allow for little or no input related to factors other than basic information about the transportation segments (such as the segments’ surfaces) or elementary geospatial raster data (such as elevation profiles). Furthermore, the focus overwhelmingly lies within the maximization of temporal efficiencies of the generated paths. If someone were to use these frameworks for the generation of personalized paths (such as ones more relatable to the concepts of walkability), they would have to leverage these primitive parameters based on their indirect correlation with the desired characteristics.

Open-Source Path-Finding

Open‑source initiatives have played a significant role in the path-finding domain, and pushed its development forward, often by leveraging the vast amounts of freely available geospatial datasets, such as those provided by the publicly-maintained OpenStreetMap project. Popular frameworks such as Valhalla (Valhalla Contributors 2025), Open Source Routing Machine (OSRM) (Luxen and Vetter 2011), and GraphHopper (Contributors 2025) employ a mixture of path-finding algorithms and address the diversity of possible requirements (such as the different modes of transportation) through customizable configurations. Such configurations frequently focus on, for instance, defining assumptions about travel speeds on specific physical surfaces. The frameworks then use these definitions to readjust the transportation networks’ cost distributions.

Both GraphHopper (Contributors 2025) and OSRM (Luxen and Vetter 2011) execute the network tuning with pre-defined “profiles”. After the adjustment, they rely on a mix of multi‑level search algorithms, graph contraction techniques, and bidirectional heuristic functions. In GraphHopper, profiles can configure mappings to specific geospatial elements and construct mode-aware cost functions and objective “weightings”. For instance, a weighting can be such that it minimizes the travel time or the travel distance. With these, GraphHopper can also be configured to consider more specific factors, such as elevation or curvature. These parameters are reflected both in the estimation of edges’ importance during the contraction hierarchies pre-processing and during the search time. For the search, and depending on the specific case, GraphHopper uses either uni-directional or bi-directional A* and Dijkstra (Contributors 2025).

An approach similar to GraphHopper is implemented by OSRM (Luxen and Vetter 2011). Like GraphHopper, OSRM uses predefined “profiles”, which define the anticipated speeds and eligibility criteria for specific segment types. Additionally, OSRM also supports setting various penalty flags for specific features within the network, such as intersections or U-turns. OSRM by default relies on a multi-level variation of the Dijkstra algorithm (derived from Delling et. al. (Delling et al. 2017; Luxen and Vetter 2011)). Using the pre-calculated matrices, OSRM then builds a multi-level partitioned graph and initiates the search. It is generally only in instances where the problem’s distance matrices grow very large that contraction hierarchies are used (Luxen and Vetter 2011).

Valhalla presents a slightly different approach by using dynamic costing functions, which are applied during the search’s runtime. Rather than relying on precomputed weights in a graph, Valhalla computes the costs of traversing through segments and junctions dynamically, while leaving the underlying segment network unmodified. The specifications of different mode profiles are, therefore, defined by the costing method and its parameters. In Valhalla, the logic of these costing methods is implemented in a similar fashion to OSRM or GraphHopper. Elements such as elevation, traffic rules, junctions, or segments’ surfaces are leveraged to define user preferences. Then, during the search phase, the logic defined by the selected costing class is injected into the usual routine of the unidirectional or bidirectional A* (Valhalla Contributors 2025).

All three of these open-source frameworks allow for some degree of preference input, either through modifying the static graphs of segments or by parameterizing the cost calculations during runtime. These frameworks are (to some limited degree) capable of reflecting sets of specific preferences in their route generation, by filtering outright ineligible options (such as highways for walking), and by rewarding or penalizing specific features of segments and junctions (such as their surfaces). Nevertheless, they require manually constructed definitions of the preferences, and are generally built upon only a fraction of the available data, ignoring much of the full geospatial context.

In contrast to these open-source frameworks, Socharoentum et. al. attempts to integrate more context into pedestrian path generation (Socharoentum and Karimi 2016). The underlying motivation lies in acknowledging that pedestrian route-finding is more than a simple shortest-path problem, but requires a careful consideration of other contextual factors as well. Therefore, they decompose each edge of the pedestrian network into data points that reflect a larger set of attributes (such as slope variation, air pollution level, and estimated shade levels) so that each pedestrian segment can be evaluated against multiple criteria and recombined into an exact composite score for route selection. Unlike the afore-discussed frameworks, they incorporate more than only the factors strictly associated with the infrastructural segments themselves, and have, in experimental results, shown an ability to reveal more pedestrian-friendly walking paths, rather than only the quickest ones. Nevertheless, it is still very limited by its dependence on static settings (which define the user preferences and the algorithm’s parameters) and a narrow selection of the considered factors.

Proprietary Route-Finding Algorithms

In terms of sheer popularity, open-source path-finding frameworks can hardly compare with the proprietary alternatives, such as Google Maps (Schröder-Bergen 2024). There is, however, generally little known about the inner workings of these services, and theories about the specific implementations and routing strategies are mostly based on speculations. Google Maps is believed to rely on a familiar mixture of contraction hierarchies, and Dijkstra and A* variants (Deshmukh 2025). Unlike the open-source services, however, Google Maps leverages immense amounts of data (including data from its proprietary crowd-sourcing streams) in graph-based machine learning models. These models are leveraged in conjunction with other methods to not only generate paths based on existing conditions (for example, historical patterns, geospatial information, or user feedback) but also to predict future events, such as possible traffic jams. Therefore, Google Maps utilizes a type of dynamic costing. Moreover, it is apparent that Google Maps uses some equivalent of the open-source routing profiles. For instance, Google Maps can optionally provide “eco-friendly” automotive routes, estimated based on types of roads, elevation profiles, or junction types and frequencies (Moondra, n.d.). Nevertheless, it is difficult to deduce much about the routing logic Google Maps implements. It is also unclear whether or not Google Maps reserves some of its tools (for instance, its graph-based machine learning algorithms) only for specific transportation modes, such as automotive transportation (Lau 2020; Deshmukh 2025).

Machine Learning-centered Approaches

The recent popularity of machine learning-based solutions has also been reflected in the path-finding domain, and several research projects have aimed to optimize or expand the capabilities of the traditional path-finding approaches.

Wang et. al. presented an approach which eliminates the need for manually crafted heuristic and cost functions by replacing them with trainable neural networks (J. Wang et al. 2019). They suggest using a Recurrent Neural Network (RNN) to model the expected cost from the initial to the currently-iterated state. The proposed RNN calculates the log-negative probability of a user taking a step, while considering a vectorized representation of the user, the current location, and the temporal context. To enhance A*’s ability to provide accurate estimates of the remaining cost (which is often calculated with primitive methods such as Euclidean distance), they suggest leveraging a moving state of the user in a value model, built upon a graph attention network, to capture the structural characteristics. The validity of this approach was demonstrated with several experiments, where raw GPS logs were matched onto a segment graph built from OpenStreetMap data, and each trip was converted into a sequence of junction visits. Vectorized representations combining unique embedding vectors and discretized temporal features (specifically, the day of the week and the hour of the day) were then assigned to each user, and at each junction, concatenated with a location embedding. These embeddings then served as an input into the RNN, and the moving‐state output fed into the value network.

For some time, there has also existed an overlap between path-finding and natural language processing. Due to their integral predispositions, LLMs do not offer only an effective way to parse and process large volumes of various textual data, but also provide a degree of “reasoning”. This reasoning can be particularly useful for evaluating various contexts against some requirements specified by, for instance, user queries.

A relatively straightforward way to leverage the potential of LLMs is to integrate them into retrieval augmented generation (RAG) systems. These systems improve the generation capabilities of LLMs by extracting relevant data from external knowledge bases and providing it as context to the LLMs, thus extending their knowledge and minimizing hallucinations (Gao et al. 2023).

Marcelyn et. al., for instance, introduce a RAG-like approach adjusted specifically for path-generation (Marcelyn et al. 2025). They aim to leverage the versatility of LLMs’ reasoning to produce personalized routes without any further training. During the initialization phase, the framework queries geospatial datasets and processes sets of trajectories of historical paths into concise pieces of human-readable descriptions of the shortest, fastest, and most frequently taken routes between two destinations. These descriptions are then encoded into embedding vectors and stored in a vector database. During inference, a user query is embedded with the same encoder and used for comparison against the stored vectorized routes, the closest few of which are retrieved and injected into a prompt passed into an LLM. The LLM is then run to consider the compiled context and returns a sequence of segments that aim to maximize the similarity to the user query. This way, they allow for a certain degree of user preference input, but rely heavily on historical data to perform route planning, and only a restricted selection of the full geospatial context.

Automated Walkability Assessments

Similarly to pedestrian path-finding frameworks, tendencies to omit the full geospatial context are also evidenced by the limited volume of research focused on exploring automated evaluation of walkability through computer science methods. Although it is apparent that existing services and tools used for automated walkability assessments (such as the NWI or WalkScore) are, overall, insufficient, there hasn’t been much academic research into exploring ways to replace them. Besides a few proposed methods (for example, published by Horak et. al. (Horak et al. 2022)) that leverage the concept of “n-minute cities”, few projects have explored assessing infrastructure and physical environments through modern computing and statistical methods (such as machine learning).

The main direction in the few existing computer science-centered projects has been leveraging vision networks, particularly on imaging data aggregated from street-view datasets. A number of published approaches employ a general workflow that systematically samples and processes street-view imagery through machine learning-based segmentation to derive quantitative indicators relevant to evaluating the environment from a pedestrian perspective (Yin and Wang 2016; Miranda et al. 2021; Z. Huang et al. 2024; X. Huang et al. 2024). Broadly, these methods collect images in regular intervals along street networks and then conduct a pixel-level analysis via various segmentation neural networks to delineate relevant environmental elements. The extracted features are subsequently aggregated into predefined indices and integrated with complementary spatial or behavioral analyses (e.g., space syntax integration values, deviations from shortest-path trajectories in GPS data, pedestrian counts, or facility accessibility) to validate various properties, including their relevance to determining walkability.

Reflecting the advancements of image-based neural networks over the past several years, these visual models are very capable of assessing the physical environment and providing insights valuable for conducting walkability assessments. Nonetheless, street-view imagery generally suffers from limited temporal representativeness, as it is infrequently and inconsistently updated. Imaging datasets necessary for covering entire cities and regions are also demanding to upkeep and aggregate. These frameworks may also often overemphasize strictly visual features. Finally, using composite indices can be susceptible to generic scoring, assuming universal preferences that are incompatible with specific regions.

Geospatial Datasets

Imaging datasets, fortunately, do not represent the only modality relevant to geospatial tasks. In fact, in recent years, large volumes of text-based geospatial datasets have been made available to the public both through publicly and privately managed initiatives. Furthermore, several of these data sources are spatially universal and aim to record data from the entire world.

Centrally-managed Datasets

Perhaps unexpectedly, several such data sources have been enabled by major private data collectors. Overture Maps (Overture Maps Foundation 2025) is one such example, and has received much attention for its public mission (Sawers 2024). It was founded to oppose Google Maps’ dominance in geographic data aggregation, and has since relied on data contributed by its core members and external contributors, such as Meta, Microsoft, AWS, or Tripadvisor. Furthermore, Overture Maps has also extracted data from governmental and open datasets. The aggregated dataset contains information about several feature types, including buildings, geographical boundaries, and various points of interest.

In contrast, the Foursquare dataset (Foursquare Labs, Inc. 2025) is focused exclusively on commercial places of interest, such as establishments or offices. Furthermore, Foursquare aggregates its data from a diverse variety of sources (Foursquare 2021). Large amounts of data come from open web crawling, during which Foursquare scrapes directory websites or chain locator pages. Some data is also extracted from third-party licensors, other data representatives, and datasets curated by various communities or outsourced teams.

Nonetheless, based on a manual review conducted as part of this work (limited to the boundaries of Cambridge, UK), both Overture Maps and Foursquare exhibit vast temporal inaccuracy and undesirable data filtering. Both datasets include numerous entries that no longer exist (such as closed establishments or temporary food tents), and omit points important for estimating walkability (such as public amenities or aesthetic-inducing and accessibility elements).

Crowd-sourced Datasets

However, in addition to centrally managed datasets, the geospatial data domain has also long involved crowd-sourcing projects and initiatives, which, across many regions of the world, have resulted in vast and diverse data sources. OpenStreetMap (OSM) (OpenStreetMap contributors 2017) is a central pillar of this movement.

OSM is a free, open‐content map dataset of the world, and is maintained by a global community of volunteer mappers, who gather geographic information through various field surveys (often relying on GPS tracking), aerial and satellite imagery tracing, and aggregation of other freely licensed sources. Contributions are collected into a centralized database and released under the loosely restrictive Open Database License.

OSM achieves a high versatility of the geospatial data it can support by relying on only a few simple data types. “Nodes” (for singular points), “ways” (for ordered sequences of nods), and “relations” (for groupings of nodes and ways and constructing more complex map features) are the three geometric primitives OSM leverages, and which, extended with sets of textual key-value pairs, can supply extremely high descriptiveness and spatial accuracy. Hence, OSM regularly contains information about all sorts of various urban elements (such as artworks, various infrastructure, or street furniture), which other datasets do not include or could not even support. Another unique strength of OSM is its connectivity with other online databases, such as Wikidata - a large public knowledge base that also includes entries of places, buildings, and streets.

Nevertheless, while in many parts of the world OSM provides a highly descriptive and temporally accurate coverage supported by active communities of contributors, in some regions, OSM’s coverage is far from satisfying. Herfort et. al. conclude in a recent analysis that in a considerable sample of global cities, OSM’s coverage of building footprints exceeded 80% completeness in 14% of the analyzed cities, but falls short of 20% completeness in 69% of the sample (Herfort et al. 2023). This inconsistency across cities is related to the respective geographical context, as the completeness remains relatively high in Europe and Central Asia (71%), and North America (64%), but is considerably lower in Latin America & the Caribbean (20%), East Asia & Pacific (20%), Middle East and North Africa (12%), and South Asia (9%). OSM’s daily contribution statistics (monitored in the online OSMstats project (Neis 2025)) only underline these trends. In the long term, European and North American countries (such as Germany, the United States, France, and the United Kingdom) are associated with significantly larger cohorts of individual contributors and receive disproportionately higher numbers of contributions.

Nonetheless, disproportionately better coverage of Western countries is a global problem in the mapping domain, equally concerning corporate datasets (and their dependents - such as Overture Maps (Overture Maps Foundation 2025)) (Tockner et al. 2025).

Local Administration Datasets

Lastly, some governmental offices and national agencies aggregate and provide public access to various geospatial datasets. These datasets tend to be more specialized and frequently tied to specific regions. To address the importance of aesthetics and nature underlined by the walkability-related urbanist literature, two geospatial feature types were particularly relevant in the context of our work: greenery elements and public green spaces.

While the level of greenery in a physical environment can be estimated based on a variety of factors, trees and large shrubs were assumed to be of particular importance, partly because of the climate and urbanist character in our experimental region of interest - cities in the UK. To gather information about trees, moreover, several official agency-provided sources were available. For instance, the Cambridge City Council regularly releases a carefully curated collection of trees in Cambridge (Council 2024). However, a more comprehensive picture of the tree cover can be created based on a LIDAR Vegetation Object Model (VOM) published by the Department of Food, Environment & Rural Affairs Data Services Forum (Agency 2023). VOM is a raster dataset aimed at identifying existing riparian tree cover and future opportunities for tree planting. It was generated through modeling of the Environment Agency’s 1m elevation LIDAR data, with the classification of vegetation and non-vegetation objects supported by spatial proximity analysis using various mappings and Sentinel-2 imagery.

Another relevant source is the Open Greenspace dataset, released by the Ordinance Survey (UK’s national mapping agency) (Survey 2025). The aim of Open Greenspace is to improve the accessibility and awareness of areas, such as parks and sports facilities, which are likely available to the public. Open Greenspace is updated regularly by the Ordinance Survey every six months.

Data Integration in Geographic Information Systems

The growing abundance of publicly accessible geospatial data has been leveraged by various data-centered frameworks, particularly in the realm of natural language processing.

Of particular interest has been the augmentation of LLMs with cartographic and spatial data to create interactive mapping systems and spatial intelligence frameworks. Various research projects focused on steering LLMs toward spatial reasoning by translating geospatial data (represented in different mixtures of map features, points of interest, road networks, and street‐level imagery) into text or low-dimensional embeddings (Unlu 2023; Manvi et al. 2023; Balsebre et al. 2023; Balsebre, Huang, and Cong 2024; Li et al. 2024). Generally, these works have applied some form of fine-tuning or contrastive multi-modal training to align their models’ internal knowledge with the geospatial data. The aim has been to enable the LLMs to learn to answer questions about urban areas, plan itineraries, or infer things like traffic speeds or population densities. By layering various models and tools, these architectures have been able to draw on both visual and textual cues to extract and reason about geographic information without having to construct entirely new spatial systems.

Nonetheless, despite its potential, the use of vectorized geospatial representations for other purposes has been sparsely explored. One of the few exceptions is presented by Tempelmeier et. al., who build a comprehensive dataset of geospatial entity embeddings to minimize the challenges of using OSM data in machine learning applications due to its vast scale, inconsistent tagging, and lack of structured semantics (Tempelmeier, Gottschalk, and Demidova 2021). They address these issues by generating embeddings of OSM entities with their own encoder, capturing spatial location and semantic representation of the available data. This approach’s usefulness is later demonstrated in an assertion task (where the embeddings were leveraged to assign Wikidata classes to the OSM entities), and a link prediction task (where countries of origin of OSM entities were predicted).

References

  • Environment Agency. (2023). Lidar Vegetation Object Model (VOM). Defra Data Services Platform. Department for Environment, Food & Rural Affairs. https://environment.data.gov.uk/dataset/ecae3bef-1e1d-4051-887b-9dc613c928ec
  • Balsebre, P., Huang, W., & Cong, G. (2024). LAMP: A Language Model on the Map. arXiv Preprint arXiv:2403.09059.
  • Balsebre, P., Huang, W., Cong, G., & Li, Y. (2023). Cityfm: City Foundation Models to Solve Urban Challenges. arXiv Preprint arXiv:2310.00583.
  • GraphHopper Contributors. (2025). GraphHopper Routing Engine. https://github.com/graphhopper/graphhopper
  • Cambridge City Council. (2024). Tree Data. https://www.cambridge.gov.uk/tree-data
  • Delling, D., Goldberg, A. V., Pajor, T., & Werneck, R. F. (2017). Customizable Route Planning in Road Networks. Transportation Science, 51(2), 566–591.
  • Deshmukh, A. (2025). From A to B: Algorithms That Power Google Maps Navigation. Towards AI. https://towardsai.net/p/artificial-intelligence/from-a-to-b-algorithms-that-power-google-maps-navigation
  • Foursquare. (2021). How to Get POI Data Right. Foursquare Blog. https://location.foursquare.com/resources/blog/leadership/how-to-get-poi-data-right
  • Foursquare Labs, Inc. (2025). Foursquare Open Source Places Dataset. https://opensource.foursquare.com/os-places/
  • Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., & Wang, H. (2023). Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv Preprint arXiv:2312.10997, 2, 1.
  • Herfort, B., Lautenbach, S., de Albuquerque, J. P., Anderson, J., & Zipf, A. (2023). A Spatio-Temporal Analysis Investigating Completeness and Inequalities of Global Urban Building Data in OpenStreetMap. Nature Communications, 14(1), 3985.
  • Horak, J., Kukuliac, P., Maresova, P., Orlikova, L., & Kolodziej, O. (2022). Spatial Pattern of the Walkability Index, Walk Score and Walk Score Modification for Elderly. ISPRS International Journal of Geo-Information, 11(5), 279.
  • Huang, X., Zeng, L., Liang, H., Li, D., Yang, X., & Zhang, B. (2024). Comprehensive Walkability Assessment of Urban Pedestrian Environments Using Big Data and Deep Learning Techniques. Scientific Reports, 14(1), 26993.
  • Huang, Z., Wang, B., Luo, S., Wang, M., Miao, J., & Jia, Q. (2024). Integrating Streetscape Images, Machine Learning, and Space Syntax to Enhance Walkability: A Case Study of Seongbuk District, Seoul. Land, 13(10), 1591.
  • Lau, J. (2020). Google Maps 101: How AI Helps Predict Traffic and Determine Routes. Google. https://blog.google/products/maps/google-maps-101-how-ai-helps-predict-traffic-and-determine-routes
  • Li, Z., Xu, J., Wang, S., Wu, Y., & Li, H. (2024). StreetviewLLM: Extracting Geographic Information Using a Chain-of-Thought Multimodal Large Language Model. arXiv Preprint arXiv:2411.14476.
  • Luxen, D., & Vetter, C. (2011). Real-Time Routing with OpenStreetMap Data. In GIS ‘11: Proceedings of the 19th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (pp. 513–516). https://doi.org/10.1145/2093973.2094062
  • Manvi, R., Khanna, S., Mai, G., Burke, M., Lobell, D., & Ermon, S. (2023). GeoLLM: Extracting Geospatial Knowledge from Large Language Models. arXiv Preprint arXiv:2310.06213.
  • Marcelyn, S. C., Gao, Y., Zhang, Y., Gao, X., & Chen, G. (2025). PathGPT: Leveraging Large Language Models for Personalized Route Generation. arXiv Preprint arXiv:2504.05846.
  • Miranda, A. S., Fan, Z., Duarte, F., & Ratti, C. (2021). Desirable Streets: Using Deviations in Pedestrian Trajectories to Measure the Value of the Built Environment. Computers, Environment and Urban Systems, 86, 101563.
  • Moondra, M. (n.d.). Navigate More Sustainably and Optimize for Fuel Savings with Eco-Friendly Routing – Google Maps Platform. Google. https://mapsplatform.google.com/resources/blog/navigate-more-sustainably-and-optimize-fuel-savings-eco-friendly-routing/
  • Neis, P. (2025). Statistics of the Free Wiki World Map. OSMstats. https://osmstats.neis-one.org/
  • OpenStreetMap Contributors. (2017). Planet dump retrieved from https://planet.osm.org. https://www.openstreetmap.org
  • Overture Maps Foundation. (2025). Overture Maps Foundation. https://overturemaps.org
  • Sawers, P. (2024). Backed by Microsoft, AWS and Meta, the Overture Maps Foundation Launches Its First Open Map Datasets. TechCrunch. https://techcrunch.com/2024/07/24/backed-by-microsoft-aws-and-meta-the-overture-maps-foundation-launches-first-open-map-datasets/
  • Schröder-Bergen, S. (2024). Google Maps and OpenStreetMap: The Coexistence of Two Unequal Siblings. Geoawesome. https://geoawesome.com/google-maps-and-openstreetmap-the-coexistence-of-two-unequal-siblings
  • Socharoentum, M., & Karimi, H. A. (2016). Multi-Modal Transportation with Multi-Criteria Walking (MMT-MCW): Personalized Route Recommender. Computers, Environment and Urban Systems, 55, 44–54.
  • Ordinance Survey. (2025). OS Open Greenspace. OS Download Products’ Documentation. https://docs.os.uk/os-downloads/topography/os-open-greenspace
  • Tempelmeier, N., Gottschalk, S., & Demidova, E. (2021). GeoVectors: A Linked Open Corpus of OpenStreetMap Embeddings on World Scale. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, 4604–4612.
  • Tockner, L., Herfort, B., Lautenbach, S., & Zipf, A. (2025). Corporate Mapping in OpenStreetMap – Shifting Trends in Global Evolution and Small-Scale Effects. Geo-Spatial Information Science, 1–20.
  • Unlu, E. (2023). Chatmap: Large Language Model Interaction with Cartographic Data. arXiv Preprint arXiv:2310.01429.
  • Valhalla Contributors. (2025). Valhalla: Open Source Routing Engine for OpenStreetMap. https://github.com/valhalla/valhalla
  • Wang, J., Wu, N., Zhao, W. X., Peng, F., & Lin, X. (2019). Empowering A Search Algorithms with Neural Networks for Personalized Route Recommendation*. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 539–547.
  • Yin, L., & Wang, Z. (2016). Measuring Visual Enclosure for Street Walkability: Using Machine Learning Algorithms and Google Street View Imagery. Applied Geography, 76, 147–153. https://doi.org/10.1016/j.apgeog.2016.09.02.
← Previous: Walkability Chapter 2: Background
Next: Walkability Chapter 4: Design and Implementation →