Stampede Predictor is a real-time computer vision system built during Hackhazards ’25 to identify dangerous crowd density and movement patterns from live or recorded video. The goal was to detect potential stampede risks early enough for organizers and public safety teams to take action. What the system does: - Processes live camera feeds or uploaded videos in real time - Detects people using YOLOv8 and maps them onto spatial grid regions - Computes crowd density and classifies risk levels from normal to critical - Streams live risk updates to the UI using Server Sent Events - Triggers audio alerts when critical risk thresholds are crossed - Produces downloadable output videos with visual overlays My contributions: - Designed the Flask backend architecture for real time video processing - Implemented YOLO based detection and spatial risk grid logic - Integrated Fluvio for event streaming and real time data propagation - Implemented Server Sent Events for live UI updates - Built the audio alert trigger system for critical risk scenarios - Structured the system for deployment readiness and modular extension Technical focus: A key challenge was maintaining low latency while processing video frames, computing risk metrics, and updating the UI in real time. This was addressed through efficient frame handling, threading and queue based pipelines, and decoupled streaming using Fluvio. Tech stack: Python, Flask, OpenCV YOLOv8 Nano (Ultralytics) Fluvio - Event streaming HTML, CSS, JavaScript Server Sent Events, threading, queues Outcome: Top 15 project in the Fluvio Track at Hackhazards ’25, selected from over 17,000 participants across 25+ countries. GitHub: https://github.com/gabsgj/Stampede-Predictor Demo: https://youtu.be/KKmF_QUh2yI Devfolio: https://devfolio.co/projects/stampede-predictor Team Arete: Gabriel James, Nayana Shaji, Jany Sabarinath, Vrindha P #ComputerVision #RealTimeSystems #Flask #YOLOv8 #CrowdSafety #EventStreaming #Python #SoftwareEngineering