All work02 / 04

SDSU STAR Lab · Graduate Research Assistant

Computer Vision for Traffic Safety

What can drone footage of an intersection say about how close vehicles come to a collision?

Schematic illustration · not project data

Role
Graduate Research Assistant
Where
SDSU STAR Lab, San Diego State University
Period
Jun 2026 – Present
Context
Caltrans-funded intersection safety study
Methods
YOLO-based detection, multi-object tracking, time-to-collision, post-encroachment time

Overview

From a video of an intersection to measurements of risk.

Crash records describe what already went wrong. Surrogate safety measures describe how close road users come to a conflict, which makes it possible to study an intersection before crashes accumulate. Computing them needs the path of every vehicle through the scene.

I build the part of the study that produces those paths from aerial drone footage, and the measures that are computed from them.

Pipeline

  1. 01

    Detection and tracking

    YOLO-based vehicle detection and multi-object tracking across the frames of each recording, producing one trajectory per vehicle.

  2. 02

    Safety measures

    Time-to-collision and post-encroachment time are computed from the extracted trajectories. I also analyze the cases where the two measures disagree.

  3. 03

    Field recording

    I work the drone recording sessions at the study's field sites in Southern California and own the footage from there on.

  4. 04

    Input validation

    Footage is prepared and validated before processing, with checks that catch unusable input before it reaches the pipeline.

  5. 05

    Reproducibility

    The pipeline and its data handling are documented so other lab members can rerun it on new footage.

Status

The study is ongoing. The illustration on this page is a schematic and does not show study data.