InitialIniIntiaInitial Problem Statement

Intersection-level traffic flow information is foundational for designing and evaluating any transportation policy meant to reduce congestion, improve traffic flow, and improve safety for pedestrians and non-motorized vehicles. Today, understanding how traffic moves through the city’s intersections relies on manual counts and custom, one-off, short-duration video and sensor analysis at selected sites. This leaves data coverage sparse in both space and time, limiting our visibility into how changes affect the way pedestrians, motorists, and non-motorized vehicles use the city’s thoroughfares.

This project aims to help the city improve traffic conditions for pedestrians and motorized and non-motorized vehicles by turning footage from intersection traffic cameras across Boston into actionable, fine-grained, high-frequency data. This modular data infrastructure allows multiple city agencies to carry out evidence-based design, evaluation, and monitoring of transportation policies and improve outcomes for city residents.

Post-Summer Report

The Problem

Boston’s Department of Transportation (DoT) needs reliable intersection-level mobility data (pedestrian, bicycle, and vehicle counts and movements) to identify hazardous locations, retime signals, and evaluate safety policy. Currently, they rely on one-time, costly third-party traffic studies that only cover a handful of intersections, for a few hours at a time, on specific days. At the same time, the city has approximately 900 signalized intersections with traffic cameras that are primarily used by traffic engineers for real-time monitoring. The video footage from these traffic cameras can only be stored for 60 days before being deleted.  Department of Innovation and Technology is interested in extracting data from this existing footage to generate a Data Foundational Layer that supports the requirements from different stakeholders in the Department of Transportation (DoT) to enable them to analyze, design, evaluate, and improve current traffic policies.

Social Impact

This project aims to improve the lives of Boston’s transportation users by providing the city with consistent, long-term, fine-grained, intersection-level mobility data. This data would enable the city to identify safety issues, monitor changes in pedestrian and vehicle activity, optimize traffic signals, evaluate transportation policies, and direct limited resources to where they are most needed. By reducing the cost and effort of data collection, the project strengthens Boston’s capacity to make evidence-based decisions that lead to safer, more efficient, and more equitable streets.

Approach

We built a modular computer vision pipeline that converts raw DoT camera footage into structured, reusable mobility data, enabling the City of Boston to establish a Data Foundational Layer (DFL). The DFL serves as a centralized resource that allows different DoT stakeholders — such as traffic engineers, Vision Zero, and curbside management teams — to evaluate, analyze, and audit current traffic policies using consistent, intersection-level data.

The DFL is designed to be extensible: new traffic metrics can be added without reprocessing existing footage, as they are derived directly from the base data already extracted. Current tracked events include pedestrian crossings, vehicle turns, and vehicle through-movements, from which a wide range of analyses and metrics can be generated.

Major Findings & Developments: A Reusable Video-to-Mobility Data Foundation

The project’s central deliverable is a durable data infrastructure:

• An 8-stage, database-backed pipeline: raw footage → camera-motion slicing → object detection → tracking → trajectories → road-region segmentation → event counts (crossings, turns) → aggregated 15-minute counts in CSV files. Each stage writes its output to PostgreSQL.
• Database-decoupled, reusable stages: because intermediate outputs (detections, tracks, trajectories, segmentation polygons) are stored as structured records, expensive steps (GPU-based detection and segmentation) never need to be rerun to test a new tracker, threshold, or event definition. New questions can be answered by querying existing data.
• Built-in reproducibility: every run stores its model name, parameters, and configuration alongside its output.
• A camera-motion detection stage, needed because Boston’s cameras are periodically repositioned by traffic engineers; it splits footage into stable “slices” so downstream analysis only touches usable, geometrically consistent video.
• The City can later add new traffic metrics (bike-lane encroachment, curb dwell time, near-miss detection, congestion) by writing new queries against already-extracted detections and trajectories to avoid manual video review and repeated GPU processing costs.
• Component performance snapshot (secondary): detection/tracking models and pedestrian/turn-counting methods were selected and tuned (e.g., YOLOv8x detector, botsort_lite tracker), with turn-count error around 13% and pedestrian-crossing error around 4–10% on a small validated sample (4–6 intersections). The preliminary results show promise in capturing similar count metrics to the status quo third-party methods, but have not yet been tested outside of these limited samples.

Key Recommendations

1. Fund a broader validation phase before wide deployment: expand ground-truth comparisons beyond the current 6 intersections and stratify by weather, lighting, occlusion, and camera type.
2. Start with a small set of high-quality, stable cameras to prove reliability before expanding across the ~900-intersection network.
3. Keep humans in the loop: manual review is currently required to align camera perspective with street-referenced directions; an automated heading-estimation prototype exists but isn’t accurate enough yet.
4. Use outputs as a decision-support foundation, not for enforcement or individual identification. Develop policies and regulations to govern access to raw video and fine-grained trajectory data accordingly.