Project overview

  • This project explores how observed behavior can become a design layer for urban design. It uses streetscape video to map how people and vehicles move through public space, with the goal of turning those patterns into evidence that can inform future design decisions. It is an early pilot and remains a work in progress.
  • The pilot compares three 10-minute videos showing Times Square, New York City, one each from summer, spring, and winter. The same three sample windows were analyzed in each video. Together, they suggest a redistribution of activity rather than a simple “more or less activity” story: summer and spring align on pedestrian coverage and edge movement, while winter diverges on both, and roadway coverage varies substantially across all three.
  • The opportunity is a design-feedback system. With standardized capture and calibration, the method could support site diagnosis, before-and-after evaluation, and future pedestrian–micromobility coexistence studies.

Pilot scope: three 30-second windows from each recording. The findings describe camera-visible activity and do not estimate total attendance or a causal seasonal effect.

Validated comparisons

Median camera-visible activity across the three sample windows for Summer, Spring, and Winter.

Each row uses its own scale, and camera geometry differs across recordings. Compare patterns within a row rather than total activity across camera views.

Go to Activity atlas →

Measured differences

Pedestrian circulation
occupied share
Summer 40.0%
Spring 44.4%
Winter 25.4%
Summer–winter difference Summer +12.4 percentage points
Roadway
occupied share
Summer 20.8%
Spring 60.0%
Winter 39.5%
Summer–winter difference Winter +20.8 percentage points
Edge / obstacle
movement intensity
Summer 0.000321
Spring 0.000324
Winter 0.000441
Summer–winter difference Winter +0.000120 normalized optical flow

Pattern across three seasons

No statistical correlation can be established from three recordings. The clearest recurring relationship is that Summer and Spring are close in pedestrian coverage and edge movement, while Winter shows lower pedestrian coverage and higher edge movement. Roadway coverage does not follow the same relationship: Spring is highest, Winter is intermediate, and Summer is lowest.

Heatmaps

Choose a zone and season. The interpretation beside the heatmap applies only to the selected zone in that recording. Brighter areas show where movement was concentrated during the middle sample window.

Summer

Summer in pedestrian circulation

Across the three sampled windows, Summer activity occupied 40.0% of the circulation zone, close to Spring and broader than Winter.

Pattern across three seasons Summer and Spring align; Winter is lower. Interpret the heatmap within its camera view; geometry, visible footprints, and heatmap scaling differ across recordings.
Validated scene-level activity measure

Manual QA frames

Each contact sheet shows ten frames sampled across the selected 30-second window. These are the source images used to visually audit detection quality and decide which measurements were safe to report.

Implications

The value of the prototype is the possibility of making observed behavior a repeatable design input.

Behavior becomes a design layer

Movement concentration can sit alongside sun, wind, access, and circulation studies. It gives teams a visual record of where a space attracts, channels, or displaces activity.

Interventions become testable

Using the same camera and measurement zones before and after a change could show whether seating, barriers, crossings, or programming altered how space was used.

The method can extend to mobility

A calibrated version could examine how pedestrians, bicycles, scooters, delivery devices, and other small mobility modes negotiate shared public space.

Next steps

Run a controlled repeat study from one fixed camera at comparable times and conditions, add ground-plane calibration, and define one design question in advance. That would move the work from a promising observational prototype toward evidence that can support a design decision.

How to read the evidence

This pilot measures spatial activity, not individual identity. The reliable outputs are deliberately limited to what the videos can support.

Space coverage
The share of small analysis cells within a zone that showed visible activity. It describes how broadly activity spread through the zone; it is not a people count.
Movement intensity
A pixel-level measure of how much the image changed between frames. It helps locate active areas, but it is not walking speed or distance.
Heatmaps
Brighter areas indicate stronger or more frequent movement within that recording. Camera geometry, visible footprints, and heatmap scaling differ, so interpret each map within its own view.
Cross-season interpretation
The report compares recurring camera-visible patterns across Summer, Spring, and Winter. With only three recordings and different camera views, it does not claim a statistical correlation or causal seasonal effect.

What the pilot supports

Zone-level patterns of movement coverage and intensity within each camera view, plus three repeatable summer–winter proxy differences.

What it does not claim

Reliable person counts, dwell behavior, hesitation, near misses, metric speed, total Times Square attendance, or a causal seasonal effect.

The next question

Can a standardized capture protocol distinguish the effect of a design intervention from the effects of camera position, event programming, weather, and crowd conditions?

Activity atlas

Compare time-normalized image-space activity within camera-specific semantic zones. Summer and winter retain their paired shared scale. Spring is shown separately with its own camera-specific scale.

Suppressed exploratory evidence
Person and vehicle presence, flow, and dwell maps are retained for auditability, but their detection and tracking QA did not support quantitative behavioral interpretation. They must not be used as person counts, unique-person estimates, dwell findings, hesitation findings, or micromobility observations.
Summer
Spring
Winter
Paired
99th-percentile scale
HighLow

Shared summer–winter scale; different camera geometry and visible footprint.

Selected window Crowd-level image-space proxy
Summer (event episode) —
Winter (snow episode) —
Summer − winter —
This selected map is suppressed exploratory evidence. No numerical behavioral interpretation is shown.

Spring

The spring recording uses the same three timestamp windows and corrected processing rules, but a different nighttime camera view. Reliable findings are limited to within-camera motion and occupied-grid proxies.

0.000811 Middle-window event-core optical flow — the spring peak.
57.5% Late event-core occupied share, down from 72.5% early and middle.
0.001217 Late roadway optical flow, about 76% above the middle window.

Interpretive boundary

This is a nighttime spring crowd episode, not an estimate of a spring seasonal effect. Camera geometry, lighting, event programming, barriers, crowd density, and weather differ from the other recordings.

Person detection failed quantitative visual QA. Unique-person, dwell, hesitation, route-directness, gate, micromobility, and near-miss claims remain suppressed.

Spring camera-level three-window medians and ranges.
Zone Occupied median Occupied range Optical-flow median Optical-flow range
Pedestrian circulation 44.4% 44.4%–47.2% 0.000465 0.000438–0.000567
Crosswalk / interface 80.6% 64.5%–80.6% 0.000682 0.000445–0.000799
Roadway 60.0% 60.0%–68.0% 0.000697 0.000691–0.001217
Edge / obstacle 32.3% 25.8%–35.5% 0.000324 0.000294–0.000378
Spring event core 72.5% 57.5%–72.5% 0.000648 0.000471–0.000811

Spring event-core motion map

One camera-level map is decoded at a time. The animated flag screen and upper billboards are masked so their changing pixels do not become crowd activity.

Summer–winter three-window scorecard

The original paired comparison remains unchanged. Claimable rows appear first; spring is reported separately above.
Edge / obstacle Optical-flow activity 0.000321 0.000441 -0.000120 3 / 3 Claimable
Pedestrian circulation Occupied-grid share 40.0% 25.4% +12.4 pp 3 / 3 Claimable
Roadway Occupied-grid share 20.8% 39.5% -20.8 pp 3 / 3 Claimable
Crosswalk / interface Occupied-grid share 45.5% 53.6% -8.1 pp 2 / 3 Descriptive only
Crosswalk / interface Optical-flow activity 0.000476 0.000425 -0.000014 2 / 3 Descriptive only
Edge / obstacle Occupied-grid share 30.8% 34.6% -3.8 pp 3 / 3 Descriptive only
Pedestrian circulation Optical-flow activity 0.000416 0.000347 +0.000058 2 / 3 Descriptive only
Roadway Optical-flow activity 0.000219 0.000302 -0.000090 3 / 3 Descriptive only
Crosswalk / interface Accepted tracks per minute 10.000 80.000 -76.000 3 / 3 Suppressed
Crosswalk / interface Dwell events per minute 4.000 44.000 -44.000 3 / 3 Suppressed
Crosswalk / interface Gate crossings: crosswalk pedestrian gate 0.000 0.000 +0.000 2 / 3 Suppressed
Crosswalk / interface Hesitation share 54.9% 58.3% -3.7 pp 2 / 3 Suppressed
Crosswalk / interface Mean visible detections 2.144 16.140 -13.996 3 / 3 Suppressed
Crosswalk / interface Median dwell seconds 4.90 s 7.45 s -2.550 3 / 3 Suppressed
Crosswalk / interface Moving share 22.3% 21.7% +2.7 pp 2 / 3 Suppressed
Crosswalk / interface Route directness 0.750 0.598 +0.152 2 / 3 Suppressed
Crosswalk / interface Stationary share 77.7% 78.3% -2.7 pp 2 / 3 Suppressed
Pedestrian circulation Accepted tracks per minute 2.000 130.000 -126.000 3 / 3 Suppressed
Pedestrian circulation Gate crossings: pedestrian circulation gate 0.000 2.000 -2.000 2 / 3 Suppressed
Pedestrian circulation Mean visible detections 2.273 25.160 -22.780 3 / 3 Suppressed
Roadway Accepted tracks per minute 32.000 28.000 +14.000 2 / 3 Suppressed
Roadway Dwell events per minute 4.000 8.000 -4.000 2 / 3 Suppressed
Roadway Gate crossings: roadway vehicle gate 0.000 0.000 +0.000 3 / 3 Suppressed
Roadway Hesitation share 27.9% 55.6% -29.4 pp 3 / 3 Suppressed
Roadway Mean visible detections 4.173 2.676 +1.870 2 / 3 Suppressed
Roadway Median dwell seconds 4.50 s 4.80 s -4.100 2 / 3 Suppressed
Roadway Moving share 66.7% 19.3% +51.4 pp 3 / 3 Suppressed
Roadway Route directness 0.866 0.701 +0.192 2 / 3 Suppressed
Roadway Stationary share 33.3% 80.7% -51.4 pp 3 / 3 Suppressed
No scorecard rows match the current filters.
26 additional rows are suppressed or below the repeatability floor. Suppressed values remain visible only to document the failed analytical pathways.

Episode-specific crowd cores

Summer core

The dense central audience is measured as crowd-level image activity. It is excluded from individual-person and trajectory claims.

Window Occupied-grid share Optical flow
Early · 01:00–01:30 8.7% 0.000140
Middle · 05:00–05:30 4.3% 0.000130
Late · 09:00–09:30 8.7% 0.000126

Spring event core

The dense left-side spring crowd is also measured at crowd level because individual detection is incomplete.

Window Occupied-grid share Optical flow
Early · 01:00–01:30 72.5% 0.000648
Middle · 05:00–05:30 72.5% 0.000811
Late · 09:00–09:30 57.5% 0.000471

Not a paired seasonal contrast

There is no equivalent winter event-core zone, and the summer and spring crowd-core polygons cover different camera footprints. These profiles describe their own episodes only.

Validation

Automated invariants passed for all three episodes. Person-detection QA did not support counts or trajectory-derived behavioral and cognitive metrics.

Summer person detectionFailed

Recall 6.71%
Median count error 92.89%

Required: at least 75% recall and no more than 20% median count error.

Winter person detectionFailed

Recall 58.92%
Median count error 40.14%

Required: at least 75% recall and no more than 20% median count error.

Spring person detectionFailed

Frames reviewed 30
Five-second person tracks 9 / 20

Systematic missed pedestrians were visible in every window; exhaustive precision and recall were not claimed.

Detection audit — 30 frames per episode
Frames were stratified across early, middle, and late windows. Summer and winter received numerical manual audits. The spring visual audit found systematic missed pedestrians by a clear margin in all windows, so person counts were suppressed without presenting pseudo-precise recall.
Track audit — five-second segments
Winter person strata met the continuity sample threshold, but failed person-detection QA still overrides those metrics. Summer lacked enough usable audited segments; spring had only nine person tracks lasting at least five seconds, below the required pool of 20.
Gate, vehicle, and micromobility disposition
No configured gate passed manual count validation. Vehicle boxes lacked passing class-specific QA. Bicycles, scooters, and other micromobility are reported as no reliable observation; the legacy aspect-ratio wheelchair classification is disabled.
Automated invariants
All three runs passed timestamp monotonicity and 0.2-second spacing, fixed track classes, one summary per unique track, observation completeness, annotated-frame counts, finite nonnegative time-normalized heatmaps, and manifest checksum verification.

Manual QA contact sheets

Use the completed review sheets to inspect scene scale, crowding, detections, and zone context.

Summer middle-window manual QA contact sheet

Methods and provenance

The report embeds the three final corrected run manifests and their canonical evidence files. Source videos are identified by filename and hash but are not embedded.

Paired sampling

  • Windows: 01:00–01:30, 05:00–05:30, and 09:00–09:30.
  • Timestamp decoding: constant 5 FPS with five seconds of tracking padding.
  • Ground point: bounding-box bottom center.
  • Tracking reset between windows; duration calculated from timestamps.
  • No metric speed, distance, density per square meter, or near-miss measures.

Semantic zones

  • Pedestrian circulation
  • Crosswalk / interface
  • Roadway
  • Edge / obstacle
  • Summer core (summer-only crowd analysis)
  • Spring event core (spring-only crowd analysis)

Interpretive boundary

Results describe activity visible within each camera’s validated semantic zones. Weather, time of day, event programming, barriers, crowd density, and camera geometry remain major confounds.

Summer final run

Winter final run

Spring final run

Embedded data exports