Skip to content

Sample project

Street Scene Analysis

Measuring the physical character of streets at scale using semantic segmentation, object detection and colour analysis on street-level imagery.

Network analysed
310 km of streets
Sample points
6,200 locations
Images processed
24,800 (4 headings)
Indicators
7 per segment
Duration
5 weeks
Network analysed
310 km

Sampled every 50 metres

Images processed
24,800

Four headings per sample point

Indicators per segment
7

Segmentation, detection and colour

Validation agreement
±3.2 pp

Against a 200-point manual sample

Problem

What needed answering

An urban research group wanted to compare the physical quality of streets across a city: how green they are, how enclosed they feel, how much of the frontage is active, how dominated they are by vehicles.

These qualities are normally assessed by sending people out to score streets by hand. That approach produces good judgement on a small sample and becomes unaffordable past a few dozen streets, while also introducing inconsistency between assessors.

The team needed indicators that were consistent across the whole network, reproducible when re-run, and transparent enough that a planning committee could challenge any individual number and see how it was derived.

Data sources

What went in

  • Street-level imagery

    Licensed panoramic imagery sampled at fixed intervals along the network, four headings per point.

  • Road network geometry

    Open mapping data defining the segments along which points were sampled.

  • Segmentation model

    An open, urban-scene segmentation model producing per-pixel class labels.

  • Detection model

    An open object detection model for vehicles, pedestrians, street furniture and frontage features.

  • Administrative boundaries

    For aggregating segment-level indicators to reportable areas.

Method

How it was done

Sample points were placed every 50 m along the network, with four camera headings per point, giving 24,800 images. Sampling by distance rather than by segment prevents long arterial roads from being under-represented relative to short residential streets.

Semantic segmentation produced per-pixel class shares for each image: sky, building, vegetation, road, sidewalk, vehicle, person and street furniture. Class shares were averaged across the four headings to give one profile per point.

The vegetation share became a green view index, which is the established measure of perceived greenery from the street. Building and sky shares together gave an enclosure ratio, a standard proxy for how contained a street feels.

Object detection ran in parallel to count vehicles, pedestrians and frontage features. Detection and segmentation answer different questions: segmentation says how much of the view is vehicle, detection says how many vehicles there are, and a street can score high on one and low on the other.

Colour analysis extracted the dominant palette per image, which distinguished areas with consistent historic materials from those with mixed modern frontage. Every indicator was aggregated to segment and district level, with the underlying per-image values retained so any number could be traced back to the images that produced it.

Processing

The pipeline

Each stage produced an artefact that the next stage consumed, so any result can be traced back to the input that created it.

  1. 01

    Street view image

    Image set

    24,800 licensed images sampled at 50 m intervals along 310 km of network, four headings per sample point.

  2. 02

    Semantic segmentation

    Class shares

    Per-pixel classification into sky, building, vegetation, road, sidewalk, vehicle, person and furniture, averaged across headings.

  3. 03

    Object detection

    Detection table

    Counts of vehicles, pedestrians, street furniture and active frontage features, with confidence thresholds applied consistently.

  4. 04

    Colour analysis

    Palette profile

    Dominant palette extraction per image, aggregated to characterise material consistency along each street segment.

  5. 05

    Urban environment analysis

    Indicator layer

    Seven indicators per segment, joined to the network and aggregated to district level, delivered as maps and a queryable dataset.

Visualization

The output

01 · Source frame
02 · Segmentation
car · 0.96car · 0.91person · 0.88storefront · 0.74
03 · Detection
  • Building31.6%
  • Sky24.1%
  • Road20.7%
  • Vegetation11.6%
  • Sidewalk8.3%
  • Vehicle2.9%
  • Person0.8%

Data

Results table

Sample project · Street segment indicators
SegmentGreen view %Enclosure ratioSky %Vehicle %Pedestrians / img
Elm Avenue31.40.8418.26.12.4
High Street8.21.6211.414.89.1
Canal Road22.70.4129.89.20.8
Park Terrace38.90.7221.63.43.7
Station Approach4.11.948.921.312.6

Findings

The key comparison

  • Park Terracehighest green view38.9
  • Elm Avenue31.4
  • Canal Road22.7
  • High Street8.2
  • Station Approachlowest4.1
Green view index: the share of the street-level view occupied by vegetation, averaged across four headings per sample point.

Result

What the analysis showed

  • Green view index across the network ranged from 4.1% to 38.9%, and the distribution followed district boundaries far more closely than the team expected. Two adjacent districts differed by more than twenty points with no change in street type.

  • Enclosure ratio and green view were only weakly related. Several streets scored well on greenery while feeling open and exposed, which matters because those two qualities are often treated as one in streetscape policy.

  • Vehicle pixel share and vehicle counts disagreed on eleven segments. In each case the segment had few but very large vehicles, which segmentation weights heavily and detection does not. Reporting both prevented a misleading conclusion.

  • Manual scoring of a 200-point validation sample agreed with the automated green view index within 3.2 percentage points on average, which was accepted as sufficient for comparative rather than absolute use.

Deliverables

What was handed over

  • Per-image class share table for all 24,800 processed images
  • Segment-level indicator layer with seven measures, joined to the road network
  • District-level aggregation tables for reporting
  • Detection counts with confidence thresholds and model version recorded
  • Colour palette profiles per segment
  • Interactive map with indicator switching, plus a method note covering models, thresholds and validation

Next step

Want something similar for your market?

Tell us the geography, the category and the decision. We will scope the equivalent project and send a fixed price before any work starts.