Street: Multiscale Urban Infrastructure

• Street is a multi-scalar urban construct defined by its spatial, visual, and socio-economic dimensions, integrating street networks, imagery, and public interactions.
• Research employs graph theory, street-level imagery, and procedural methods to yield measurable metrics like orientation entropy and spatial planarity ratios.
• Emerging studies integrate AI infrastructures with health and economic indices, advancing street simulation, reconstruction, and urban governance frameworks.

Searching arXiv for the specified STREET-related papers to ground the encyclopedia article. Street is a multi-scalar urban construct that contemporary research treats simultaneously as a spatially embedded graph, a street-segment observational unit, a street-level visual field, a reconstructable three-dimensional scene, and a public field where mobility, commerce, health, and algorithmic governance intersect. In this literature, street is not reducible to roadway centerlines or vehicular hierarchy alone: it includes intersections and segments, road surfaces and street objects, visual interfaces and semantic frontage, exposure to pollution and collision risk, exercise opportunity, economic vitality, and the situated presence of AI infrastructures (Barthelemy et al., 2024, Boeing, 2020, Wang et al., 2019, Taylor et al., 27 Oct 2025).

1. Graph-theoretic and spatial representations

Most recent work represents street networks in primal form, with nodes as physical intersections and edges as street segments. Global OSMnx-based modeling constructs these networks as primal, nonplanar, directed multigraphs that may contain self-loops and parallel edges and retain all connected components (Boeing, 2020). The review literature treats this representation as dominant for accessibility, routing, geometry, and planning questions, while noting that dual representations can capture relations among whole streets but may obscure geometry and distance (Barthelemy et al., 2024). Street is therefore analyzed as both topology and embedding: degree, cycles, clustering, centrality, orientation, length, area, grade, and planarity are all structurally relevant.

This perspective has produced globally comparable measurement systems. One study modeled 8,914 urban areas in 178 countries, simplifying more than 160 million OpenStreetMap street network nodes and over 320 million edges to a final global set of 37 million nodes and 53 million edges (Boeing, 2020). The associated indicators include average node degree, street segment length, intersection density, orientation entropy, circuity, PageRank, elevation, and grade. The review literature argues that a parsimonious indicator set is often preferable because several classical indices are largely functions of average degree in street-network contexts (Barthelemy et al., 2024).

Orientation and entropy provide a particularly explicit measure of spatial order. For a bearing distribution binned into 36 angular classes, orientation entropy is defined as

$H = -\sum_{\theta} p_{\theta}\ln p_{\theta},$

and the orientation-order indicator is

$$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$

with $H_g=\ln(4)$ for an idealized single orthogonal grid and $H_{\max}=\ln(36)$ for a uniform bearing distribution (Boeing, 2018). Across 100 cities, $o$ ranges from 0.002 in Charlotte to 0.899 in Chicago, capturing the extent to which a city follows the geometric logic of a single grid (Boeing, 2018). The same study shows significant relationships between orientation-order and circuity, dead-ends, node degree, and four-way intersections, reinforcing the view that geometric and topological order are coupled.

The literature also stresses that streets are rarely perfectly planar. The Spatial Planarity Ratio is $\phi = i_n / i_p$, where $i_n$ is the count of nonplanar intersections in the true 3D graph and $i_p$ is the count of planarized intersections in a 2D embedding (Barthelemy et al., 2024). Across 50 world cities’ drivable networks, only $\sim20\%$ are formally planar; mean $\phi \approx 0.88$, with a low around 0.54 for Moscow (Barthelemy et al., 2024). This matters because overpasses and tunnels alter degree counts, cycles, and accessibility if planarization is handled naively.

A related planning literature extends representation beyond transport function. Street context classification replaces purely vehicular classes with categories such as Alley, Park, Downtown Commercial, Neighborhood Commercial, Neighborhood Residential, Residential Throughway, Commercial Throughway, Highway, Highway Ramp, Industrial, and Downtown Residential, thereby combining transportation function with side-use context (Alhasoun et al., 2019). This broader framing is consistent with the distinction, explicit in procedural reconstruction work, between roads and streets: prior road-oriented methods often omit urban features and context, whereas street-oriented models include surfaces, intersections, lanes, and street objects as a coherent whole (Cura et al., 2018).

2. Street-level imagery, semantics, and latent structure

Street-level imagery has become a primary sensor for semantic inference. In street scene labeling, one influential formulation treats semantic segmentation as superpixel classification with explicit spatial priors. A priori superpixel CNNs preserve the superpixel in situ by masking the remainder of the image to black, allowing the network to encode absolute image coordinates and scene structure through the fully connected layers (Wang et al., 2019). Unary terms come from softmax label scores, while refinement is performed with a soft restricted MRF:

$$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$0

where the pairwise term uses conditional label co-occurrence from retrieved scenes and a “soft restriction” weight derived from inter-superpixel score differences (Wang et al., 2019). On CamVid, the full model reaches 78.1% per-pixel accuracy and 53.2% mean-class accuracy; on SIFT Flow Street it reaches 82.0% and 41.1%, respectively (Wang et al., 2019). The main gain is not raw background dominance but foreground preservation under dataset bias and reduced over-smoothing.

Other work treats street imagery as an unsupervised source of interpretable latent structure. ConvPCA first trains a convolutional autoencoder and then applies PCA to produce ordered, orthogonal latent components from both street-level images and rasterized street-network images (Law et al., 2019). In Greater London, 110,493 street-view images are encoded into a 4,096-dimensional latent vector before PCA; globally, 107,973 street-network tiles are encoded into a 640-dimensional latent vector (Law et al., 2019). The learned components are not merely dimensionality reductions: PCA 1 relates to street urbanity and building articulation, PCA 3 trades greenery versus building density, and mapped component values show strong spatial dependence, with Moran’s I $$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$1 and $$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$2 for two principal street-level components (Law et al., 2019). This suggests that street imagery contains ordered latent factors that are geographically structured and can be used for knowledge discovery rather than only recognition.

Street-view images have also been used to operationalize planning taxonomies directly. CNN-based street context classification on Google Street View achieves validation accuracies from 81.69% to 84.17% in San Francisco and from 83.16% to 87.79% in Boston, with Inception-v3 performing best in both cities (Alhasoun et al., 2019). Embedding-space visualization shows alleys, parks, highways, and ramps forming distinct neighborhoods in feature space, while class activation mapping highlights storefronts, sidewalks, signage, high-rise facades, tree canopies, and road surface as context cues (Alhasoun et al., 2019). A common misconception is that street-view models only recover transport hierarchy; these results indicate that frontage, greenery, enclosure, and land-use expression are visually encoded at useful accuracy.

3. Network generation, procedural reconstruction, and learned morphology

Street generation research spans rule-based procedural systems, context-aware image completion, discrete autoregressive graph synthesis, and direct graph representation learning. At the procedural end, StreetGen reconstructs streets from rough GIS centerlines into road surfaces, intersections, lanes, inter-lane connections, and street objects, entirely inside PostgreSQL/PostGIS (Cura et al., 2018). It uses a strong, yet simple modelling hypothesis: streets are organized around street axes, street morphology is piecewise constant, and cornerstone paths at intersections are either straight segments or circular arcs (Cura et al., 2018). Roadway surfaces are built by buffering centerlines, $$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$3, intersections are assembled from buffer intersections and arcs, and turn trajectories are generated as cubic Bézier curves (Cura et al., 2018). The system reconstructs the entire city of Paris in less than 10 minutes on 1 core, computes a single street in ~200 ms, and reduces end-to-end time to about 1 minute with parallel clusters (Cura et al., 2018).

DeepStreet recasts street generation as image completion. Each sample is a 256×256 tile representing 1,280×1,280 m, with a 48×48 masked hole to be inpainted from surrounding street networks and terrain (Fang et al., 2020). The model uses a fully convolutional encoder–decoder generator with global and local context discriminators, weighted MSE inside the hole, and an adversarial objective with $$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$4 (Fang et al., 2020). Trained on 900,000 non-overlapping tiles from Barcelona, it predicts both gridiron and irregular street networks and runs at ~0.211 seconds per tile on an NVIDIA RTX 2080 Ti GPU (Fang et al., 2020). Its limitation is equally explicit: if the hole is surrounded by large blank areas, the model often outputs blank inside the hole because boundary cues are absent (Fang et al., 2020).

A more explicitly generative formulation models complete city-scale street networks as traversable graphs. A transformer decoder predicts a 2D field of discrete indices in a sliding-window manner, with each index referencing a learned VQVAE dictionary of local street neighborhoods. The autoregressive factorization is

$$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$5

with 16 × 16 windows and sequence length 256 (Birsak et al., 2022). Using OpenStreetMap data from the largest 50 US cities, the system generates index fields as large as 256 × 256, corresponding to ~19.5 km × 19.5 km or roughly 380–400 km², then decodes them into distance fields and finally into traversable graphs via thinning and RoadTracer-style extraction (Birsak et al., 2022). The method is conditioned on street density, high-priority roads, and land-water maps because sliding windows alone lead to incoherent generation (Birsak et al., 2022).

Direct graph representation learning avoids rasterization altogether. A graph-based variational autoencoder factorizes street-network generation as $$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$6, where a transformer models node coordinates and a VGAE models adjacency conditioned on learned node embeddings (Neira et al., 2022). Trained on 39,364 global 1 km × 1 km street-network tiles with mean nodes $$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$7 and mean edges $$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$8, it generates synthetic street configurations whose topological and geometric distributions closely match real networks, with slight bias toward more degree-2 nodes and longer streets (Neira et al., 2022). This addresses a limitation of raster approaches identified in the same paper: rasterization discards topology and cannot recover complex street-network features from images alone (Neira et al., 2022).

4. Reconstructing and simulating street scenes in 3D

Street scenes pose a distinct reconstruction problem because cameras move along long and narrow trajectories, foreground objects are dynamic, and background surfaces are sparse across multiple scales. StreetSurfGS adapts Gaussian Splatting to this setting through a planar-based octree representation, segmented training, guided smoothing, and dual-step matching (Cui et al., 2024). The octree concentrates capacity in complex regions, planar Gaussians improve road and facade stability, and SAM-filtered normal smoothing avoids regularizing across object boundaries (Cui et al., 2024). On KITTI-360, the method reports 23.90 PSNR / 0.874 SSIM / 0.246 LPIPS; on Waymo, 28.67 / 0.935 / 0.279; and on MatrixCity, depth error 1.51 and Chamfer Distance 0.29, outperforming several object-centric baselines (Cui et al., 2024). Memory on Free dataset scenes falls to about 11–12 GB, versus 14–18 GB for 3DGS and 23 GB for PGSR (Cui et al., 2024).

Compression becomes essential once street scenes are represented as millions of Gaussian primitives. SparseStreet introduces node-based learnable pruning and background compression for street-scene 3DGS (Wuwu et al., 2 Jun 2026). The total objective is $$o = 1 - \left(\frac{H_O - H_g}{H_{\max} - H_g}\right)^2,$$9, with node-aware sparsity coefficients that aggressively prune background while protecting rigid, deformable, and SMPL nodes (Wuwu et al., 2 Jun 2026). On Waymo, OmniRe drops from 1.55M to 0.46M Gaussians and increases from 46.15 to 80.22 FPS; StreetGS drops from 0.87M to 0.29M Gaussians and rises from 21.60 to 57.66 FPS (Wuwu et al., 2 Jun 2026). On nuScenes, FPS reaches 435.85 and 461.17 for compressed OmniRe and StreetGS variants, respectively (Wuwu et al., 2 Jun 2026). The central design claim is specific: global pruning harms moving vehicles with limited view-time, whereas background-only compression preserves dynamics (Wuwu et al., 2 Jun 2026).

Generative street simulation now extends beyond reconstruction. Streetscapes generates long street-level videos along user-specified trajectories from map/layout control and text prompts using autoregressive video diffusion with temporal imputation (Deng et al., 2024). It is trained on 1.5M Google Street View images covering ~33 km² across Paris, London, Barcelona, and New York, and substantially outperforms InfiniCity in long-range generation, with FID 17.79 versus 108.47 on the all-steps setting (Deng et al., 2024). StyledStreets, by contrast, edits already reconstructed street scenes across seasons, weather, and camera setups while preserving geometry across seven synchronized vehicle-mounted cameras (Chen et al., 27 Mar 2025). Its hybrid embedding disentangles geometry from style, uncertainty-aware rendering filters noisy diffusion supervision, and joint pose optimization plus multi-view training improve geometry, yielding an 18% Chamfer Distance reduction versus OmniRe and +2.15 dB vehicle PSNR in reconstruction (Chen et al., 27 Mar 2025). These systems collectively suggest a shift from scene capture toward controllable, physically anchored street simulation.

5. Environment, safety, and behavior

Street is also an environmental exposure field. In a wind-tunnel experiment on a perpendicular street canyon with two rows of model trees, the presence of trees shifted pollutant dispersion from a nearly two-dimensional to a three-dimensional field, creating strong along-street and across-street heterogeneity (Neira et al., 2022). Yet the average level of pollution in the street, and thus the overall ventilation efficiency, did not show a specific trend with the density of trees (Neira et al., 2022). Using the exchange-velocity formulation

$H_g=\ln(4)$0

the study reports $H_g=\ln(4)$1 for Zero, 0.019 for Half, and 0.020 for Full tree configurations, while concentration patterns become much more uneven under denser planting (Neira et al., 2022). This directly challenges a simplistic greening narrative: local hotspots can intensify even when bulk ventilation changes only modestly.

Driver behavior responds similarly to geometric and visual street cues. A large-scale study in Milan, Amsterdam, and Dubai finds that posting lower speed limits is not sufficient to reduce driving speeds effectively (Orsi et al., 6 Jul 2025). In Milan, the matched citywide effect of a 30 km/h limit is modest: average speed falls from 29.48 km/h to 27.19 km/h and the 85th percentile from 41.97 km/h to 38.73 km/h, yielding an ATE of −2.29 km/h and −3.45 km/h, respectively (Orsi et al., 6 Jul 2025). Narrower streets, shorter segments, fewer lanes, more building pixels, and lower road and sky pixel shares are associated with lower speeds, whereas visibility/open-ness and larger sky views encourage faster driving (Orsi et al., 6 Jul 2025). In a citywide 30 km/h simulation for Milan, only 335 km of 1,959 km are predicted to achieve $H_g=\ln(4)$2 km/h with posting alone, while 796 km require design interventions and 828 km likely require major changes, enforcement, or exclusion (Orsi et al., 6 Jul 2025).

Pedestrian safety can likewise be predicted from static built-environment cues. STRIDE introduces a benchmark of 18,036 panoramic Google Street View images in Bogotá, with 557,115 annotated instances over 9,900 panoramas across 27 static infrastructure categories (González et al., 2023). A DINO-based detector combined with a collision prediction module improves pedestrian collision frequency prediction: on the held-out test set, the count-only AutoML baseline using DINO counts plus coordinates yields RMSE 13.78 and WMAE 28.55, whereas the end-to-end baseline yields RMSE 12.88 and WMAE 23.41 (González et al., 2023). Coordinates alone are informative, but object counts and object-level self-attention further reduce error, indicating that street form and street furniture encode measurable risk signals (González et al., 2023).

6. Vitality, health equity, and AI in the street

Recent work treats street as a micro-scale economic and social diagnostic unit. The Street Economic Vitality Index composes nine indicators—Shop Density, Closure Ratio, Weighted Brand Ratio, Mall Spillover Vitality, Motor Vehicle Density, Non-motor Vehicle Density, Pedestrian Presence, Green Coverage Ratio, and Shopfront Glazing Density—into three dimensions: Commercial Activity, Spatial Utilization, and Physical Environment (Zhuo et al., 10 Apr 2026). Dimension scores are entropy-weighted, then combined by TOPSIS as

$H_g=\ln(4)$3

where $H_g=\ln(4)$4 (Zhuo et al., 10 Apr 2026). In Nanjing, SEVI reveals a core–periphery gradient around Xinjiekou, distinguishes “Healthy Density” from “Hollow Density,” and identifies a Density–Competition Paradox in which Closure Ratio correlates positively with Shop Density ($H_g=\ln(4)$5) and Weighted Brand Ratio ($H_g=\ln(4)$6) (Zhuo et al., 10 Apr 2026). A dual-stage VLM–LLM brand pipeline reaches F1 = 0.821, up from 0.412 for OCR and 0.652 for VLM only (Zhuo et al., 10 Apr 2026).

Street is equally central to health equity. A triadic framework based on conceived, perceived, and lived space models street-level exercise deprivation from street networks, street-view imagery, and social media (Zhao et al., 4 Jul 2025). XGBoost performs best, with citywide CV $H_g=\ln(4)$7 and RMSE $H_g=\ln(4)$8 (Zhao et al., 4 Jul 2025). Aggregated SHAP contributions attribute 57.0% of explained variation to conceived space, 29.6% to perceived space, and 13.4% to lived space (Zhao et al., 4 Jul 2025). Counterfactual simulations show strong contextual variation: Bao’an’s full triad reaches up to 11.53%, while the abstract reports “up to 14%” increases in exercise supportiveness (Zhao et al., 4 Jul 2025). The paper explicitly notes this discrepancy and suggests it may reflect additional simulations or parameter ranges not detailed in the main text (Zhao et al., 4 Jul 2025). The important point is methodological: street deprivation is not reduced to one scalar walkability index but classified into C-only, P-only, L-only, dual, and CPL modes via thresholded SHAP aggregation (Zhao et al., 4 Jul 2025).

A different line of research emphasizes that street is also where AI becomes infrastructural and publicly consequential. Through “everyday AI observatories” conducted across five streets in Cambridge, Coventry, Edinburgh, London, and Logan, the concept of “reciprocity deficits” describes the breakdown in mutual, situated seeing between AI infrastructures and everyday publics (Taylor et al., 27 Oct 2025). The study identifies three tensions: the street as a transactional environment, the designed invisibility of AI and its publics in the street, and the stratification of street environments through statistical governance (Taylor et al., 27 Oct 2025). This reframes explainability. The issue is not only whether models are transparent, but whether sensing, notification, recourse, and public benefit are materially visible and accountable in place.

Taken together, these strands present street as a uniquely dense research object. It is a graph with measurable order and redundancy, a visual field with semantic and latent structure, a procedural and generative design substrate, a reconstructable and stylizable 3D environment, an environmental and behavioral exposure setting, an economic and health diagnostic unit, and a site where urban governance and AI meet everyday publics. A plausible implication is that no single representation is sufficient: the most consequential street research now links topology, geometry, imagery, semantics, dynamics, and situated social practice rather than treating them as separable domains.