Street Review: Participatory AI Urban Analysis

• Street Review is a systematic, multi-dimensional framework that blends qualitative user input with AI-driven image analysis to evaluate public streets.
• It integrates resident interviews, focus groups, and transformer-based segmentation to produce detailed, demographic-specific spatial heatmaps.
• The approach generates actionable insights for urban planning by identifying design enhancements and policy interventions tailored to diverse community needs.

Street Review refers to the systematic, multi-dimensional assessment of public streets using integrated qualitative, quantitative, and AI-assisted methodologies. Street Review frameworks evaluate inclusivity, accessibility, physical quality, and user experience, combining participatory data collection (e.g., resident interviews, focus groups) with machine learning on large-scale imagery. The core objective is to generate actionable, high-fidelity analytics—such as spatial heatmaps disaggregated by demographic group and criterion—that inform equitable urban planning, policy, and management. The methodology has been operationalized in Montreal, with an emphasis on co-produced labels, transformer-based image analysis, and subgroup-specific predictive mapping (Mushkani et al., 14 Aug 2025).

1. Participatory–AI Hybrid Framework

Street Review employs an explicit co-production approach blending participatory research and AI-based analysis. The workflow comprises:

1. Semi-directed resident interviews: 28 Montreal residents, oversampling under-represented groups (recent immigrants, LGBTQIA2+, elderly, etc.), provided lived-experience data using adaptive prompts focused on inclusivity, accessibility, aesthetics, and practicality.
2. Thematic coding and image labeling: Interview data underwent double coding, distilled to four evaluation criteria. Focus groups of 12 participants rated 120 curated street images using a four-point scale (1 = poor, 4 = excellent) across all criteria. Group discussion yielded consensus labels, then ranked the most/least inclusive images.
3. Image collection and AI-based feature extraction: The labeled set (15,000 images derived from 20 streets × 3 points × 250 frames) seeded model training; ~45,000 Mapillary street-view images provided the citywide prediction base.
4. Supervised multi-output regression model: An attention-based MLP, taking sequences of 12-dimensional semantic segmentation feature vectors, was trained to predict 28 targets (4 criteria × 6 demographic groups + 4 aggregate scores) using mean squared error loss.
5. Heatmap generation and subgroup analytics: Inference on geotagged frames produced detailed heatmaps per demographic and criterion, extending perceptual assessment from the labeled sample to the full urban network (Mushkani et al., 14 Aug 2025).

This methodology integrates semi-structured interviews, focus-group consensus, and transformer-based image representation, generating both qualitative and quantitative insight.

2. Data Acquisition, Labeling, and Ground Truth

Participant recruitment leveraged outreach to over 100 community groups, resulting in 28 interviewees across prioritized demographics. The process used a semi-structured guide, audio transcription, and double-coding to extract intersectional themes, which were then collapsed to four evaluation axes.

Labeling procedures involved:

• Focus-group scoring: 120 images per group (20 streets × 3 points × 2 opposing views) were scored independently, discussed, and consensus-labeled for the core criteria.
• Propagation of labels: Group scores were propagated to all 250 frames at each sampled street-point, yielding 15,000 labeled images for supervised learning.
• Geospatial extension: The trained model predicted scores for all Mapillary images citywide, producing heatmaps without further manual scoring.

The dataset reflected deliberate oversampling of marginalized user perspectives, operationalizing the co-production imperative and supplying intersectionally robust ground truth.

3. Feature Engineering and Metric Design

Physical features extracted via semantic segmentation include:

• Sidewalk width (S): computed directly from segmentation masks (meters).
• Maintenance score (M): 1–4 integer, capturing surface state, cracks, debris, and evenness.
• Greenery index (G): $$G = \frac{\text{Area of vegetation pixels}}{\text{Total sidewalk area}}$$
• Seating density (C): $$C = \frac{\text{Number of seating objects}}{\text{Sidewalk length (m)} \times 100}$$

Aggregated indices include mean maintenance ($\bar{M} = \frac{1}{N}\sum_{i=1}^N M_i$), green coverage ratio, and seating availability per 100 meters.

Model evaluation used Pearson correlation ($\rho_{X,Y}$), permutation importance ($\Delta R^2$), and spatial visualizations (street-segment heatmaps per demographic/criterion). The segmentation model (SegFormer-B5) generated per-pixel class probabilities for 9 relevant environment classes, facilitating fine-grained feature mapping (Mushkani et al., 14 Aug 2025).

4. Machine Learning Architecture and Visual Analytics

The "Street Review" module comprises:

• Input: Flattened arrays of semantics+RGB per pixel (12D per pixel; 256 × 256 grid).
• Model: An attention-based multilayer perceptron with 11 fully connected layers and 6 multi-head attention blocks explicitly designed for image-to-perception translation.
• Targets: 28 simultaneous regression outputs—one per group/criterion.
• Optimization: Mean squared error loss; validation $R^2 = 0.91$, test $R^2 = 0.89$.

Heatmaps are generated with Folium/Leaflet, supporting analytic overlays per group and evaluation axis.

Correlation analysis revealed key axes (e.g., inclusivity-accessibility, inclusivity-aesthetics) and permutation importance established the marginal value of sidewalk width, continuity, greenery, seating, and other features in perception score prediction.

5. Demographic-Specific and Spatial Patterns

Cross-group analysis demonstrated significant heterogeneity:

• Inclusivity means (scale 1–4): mobility-impaired ≈ 1.8 (lowest); elderly male ≈ 2.0; elderly female ≈ 2.1; young female ≈ 2.3; LGBTQIA2+ ≈ 2.3; young male ≈ 2.4 (highest).
• Correlations: Strong association exists between inclusivity and both accessibility (ρ = 0.51–0.55) and aesthetics (ρ = 0.54–0.64), but weak or negative association between practicality and aesthetics.
• Spatial gradients: Central districts (Ville-Marie, Outremont) scored highest; suburban/peripheral areas lower, particularly for mobility-impaired respondents.
• Example street profiles: Low-scoring streets exhibited narrow sidewalks, little to no greenery, and a lack of seating; high-scoring streets combined wide sidewalks, planting, benches, and public art.

Demographic-specific layers in the heatmaps clarify that streets may be inclusive for one group yet less so for others, reinforcing the need for disaggregated evaluation (Mushkani et al., 14 Aug 2025).

6. Implications for Planning, Equity, and Methodological Advancement

Street Review’s principal recommendations for urban policy emphasize:

• Physical design imperatives: Sidewalk continuity, width, curb cuts, and quality, given their dual effect on accessibility and inclusivity.
• Social infrastructure: Greenery and seating not as amenities, but as core elements driving perceived inclusivity and social belonging.
• Targeted intervention: Demographic-weighted heatmaps as tools for pinpointing where specific groups experience exclusion, supporting spatially precise upgrades.
• Participatory co-production: Embedding diverse cohorts in annotation, validation, and iterative refinement minimizes hidden biases and surfaces intangible, symbolic inclusivity markers.
• Scalability: The modular pipeline, open-source codebase, and labeled datasets are adaptable and scalable to other cities or data modalities (e.g., footfall sensors, social-media geotags).

The co-production process demonstrably enhances ground truth, reveals contextually significant features (e.g., multilingual signage, public art), and allows iterative calibration based on stakeholder feedback (Mushkani et al., 14 Aug 2025).

7. Conclusions and Broader Significance

Street Review establishes participatory AI pipelines as state-of-the-art for urban public space evaluation. Its integration of semi-structured elicitation, focus-group scoring, transformer-based segmentation, and attention MLP regression produces high-fidelity, demographically disaggregated maps of inclusivity.

The approach demonstrates both the existence of universal design drivers (wide, well-maintained sidewalks, seating, greenery) and group-specific needs (e.g., lighting, wayfinding, symbolic cultural inclusions). By supplying a robust, transferable, and open methodological template, Street Review provides urban planners and analysts with a data-driven, equity-focused instrument for public street management and improvement (Mushkani et al., 14 Aug 2025).