Weak Supervision · Computer Vision

Video as a Sensor for Urban Risk

Programmatic labeling meets computer vision: we generate risk labels from behavioral heuristics and weak supervision when manual annotation is too expensive, surfacing dangerous roadway events at scale.

Felipe Felix Arias Daniel Carmody Richard Sowers Jayati Singh Kevin So

University of Illinois at Urbana-Champaign

Pedestrian crossing detection

The Problem

Dangerous events are rare. Labels are expensive.

Risky roadway events (near-misses, illegal crossings, cyclist close calls) are rare in dashcam footage but critical to detect. Hand-labeling thousands of hours of video is impractical. We needed a way to generate training labels programmatically from the video signal itself.

Approach

Heuristic labeling from detection + tracking.

We identify compound objects (e.g., cyclist = person + bicycle), track them through occlusion using temporal context, and apply behavioral heuristics (proximity, velocity, trajectory) to generate probabilistic risk labels. Matrix completion weights the noisy signals into usable training data.

Cyclist Tracking

Recovering identity through occlusion.

Original cyclist detections

Baseline: drops cyclist when partially occluded.

Enhanced cyclist tracking

Ours: temporal inference recovers missed detections.

Risk Detection

Weak supervision for rare events.

High-risk scene 1
High-risk scene 2

Probabilistic risk labels generated from pose, motion, and contextual heuristics: no manual annotation required.