ClearPath

Inspiration

Crowd disasters are rarely caused by a single failure. They emerge from human behavior under extreme density, uncertainty, and cognitive overload.

Tragedies like the 2022 Itaewon Halloween crowd crush, the 2021 Astroworld Festival disaster, the 2010 Love Parade crush, and many other large-scale events reveal a consistent pattern:
people don’t fail because they lack information, they fail because too much is happening at once.

Research on crowd dynamics and disaster psychology shows that as density rises and movement destabilizes, people experience:

loss of situational awareness
delayed reaction times
instinctive herding behavior

By the time a situation is clearly labeled an “emergency,” it is often already too late.

ClearPath is driven by one question:
What if we could intervene earlier — before panic, before collapse, before injury?

What it does

ClearPath detects cognitive overload and unstable crowd movement in large events and intervenes in real time to reduce risk.

Using wall-mounted cameras, the system continuously analyzes crowd density and motion patterns. When potentially dangerous behavior is detected, ClearPath flags the affected area and triggers a coordinated intervention:

A red illuminated zone highlights the congested or unstable region that needs resolution
A surrounding green guidance zone indicates the safe direction people should move toward

This intervention is designed to be simple, intuitive, and non-verbal, cutting through noise and confusion without overwhelming people.

Rather than monitoring crowds passively, ClearPath actively guides behavior to restore stability.

How we built it

ClearPath combines real-time computer vision, crowd behavior analysis, and an interactive visualization layer.

Perception & Risk Analysis

Implemented YOLOv8 Nano for real-time person detection
Applied optical flow analysis (Lucas–Kanade) to track crowd movement dynamics
Designed four key risk metrics: crowd density, bidirectional flow, flow conflict, and stop-go waves.

Stability Logic

Danger zones are triggered only when risky patterns persist over time, preventing flicker and allowing realistic response windows.

Frontend & Simulation

Built a React-based dashboard with real-time video overlays
Implemented a tactical map view showing a bird’s-eye perspective of the event space with simulated drone deployment to intervene at danger zones
Used canvas-based rendering and detection caching to maintain real-time performance

Challenges we ran into

We initially focused on emergency signage and alerts, but realized most venues already handle clearly labeled emergencies well. The real gap lies in ambiguous, pre-emergency situations where danger is building but not yet obvious.
We attempted to integrate an Arduino-based setup, but real-time computer vision from live video exceeded its practical limits within the hackathon timeframe.
High-quality footage of real crowd instability is rare. Many public datasets are blurry, low-resolution, or poorly annotated, making validation time-consuming.

Accomplishments that we're proud of

Translating academic crowd research into a practical, visual intervention
Achieving real-time performance through careful optimization

What we learned

Pushed us to learn computer vision while making technical decisions based on real human pain points.
Effective safety systems must reduce cognitive load, not add information
Understanding user psychology is as important as model accuracy