Inspiration
In high-density industrial corridors such as warehouses, distribution hubs, and manufacturing plants workers, forklifts, and heavy machinery constantly share tight spaces. Accidents happen in split seconds, and delayed emergency response often turns manageable injuries into life-threatening situations or costly operational downtime. Traditional facility safety relies on passive camera recordings viewed after an accident has already occurred, paired with static annual training videos that fail to capture actual floor hazards. We built Hazard Lens to transform passive surveillance feeds into an active, real-time safety ecosystem that protects workers today and prevents incidents tomorrow.
What it does
Hazard Lens is an autonomous transit and logistics safety platform featuring a closed-loop dual engine:
The Reactive Engine (Real-Time Vision & Instant Dispatch): Connects to facility camera feeds overlooking transit corridors and applies real-time computer vision with pose detection to continuously monitor human movement. The moment a worker fall or impact occurs, the system flags the exact zone on a live dashboard and automatically dispatches sub-second emergency SMS alerts to response personnel.
The Proactive Engine (AI Safety Coach & Predictive Analytics): Automatically logs every near-miss and confirmed incident into a telemetry database. An integrated LLM-powered AI Safety Coach analyzes historical hazard logs to uncover recurring risk patterns, high-density collision zones, and shift-based danger trends allowing floor managers to generate hyper-localized, scenario-based safety drills derived from actual facility data.
Vision safeguards: after a fall, the camera keeps watching for 10 seconds and reports whether the worker recovered, is still moving, or is unresponsive (a possible medical emergency). Cameras also flag when they're blinded by glare, a covered lens or smoke ("VISION IMPAIRED") instead of going silent. Two cameras can cover one zone, so a blocked view doesn't mean a missed fall. No video is stored; only incident records and skeleton data leave the camera.
How we built it
Computer Vision & Pose Tracking: Built using OpenCV and Ultralytics YOLO for real-time human pose estimation and spatial camera feed analysis.
Backend Pipeline: Powered by Python and FastAPI, enabling asynchronous frame processing and sub-second webhook triggers for instant alert dispatch.
Frontend Dashboard: Developed with React and Tailwind CSS to display real-time zone mapping, active incident warnings, and historical telemetry visualization.
AI Safety Coach: Integrated using large language models to query incident telemetry, identify spatial risk clusters, and provide conversational safety recommendations.
Emergency Alert System: Integrated with automated SMS webhooks to deliver instant emergency notifications directly to floor managers' phones in milliseconds.
Fall & Post-Fall Logic: a per-person state machine (Upright -> Falling -> Down) using hip-drop speed, torso angle and height collapse, plus joint-motion analysis to tell an unresponsive worker from one who's moving, and per-camera glare and contrast checks to detect impaired vision.
Challenges we ran into
Real-Time Video Latency: Processing multi-camera video streams without frame drops or high inference latency required optimizing our pose estimation pipeline and implementing asynchronous frame buffering.
False Positive Filtering: Fine-tuning fall detection rules so the system accurately distinguishes between a worker bending down to pick up a package versus an actual impact or fall.
Spatial Telemetry Mapping: Translating 2D video coordinates from camera feeds into clear, actionable spatial zone mapping on the manager dashboard.
Falls Toward the Camera: a person lying pointed at the camera looks upright in 2D; we added height-collapse detection to catch it.
Model Jitter vs. Stillness: small frame-to-frame joint noise made motionless workers look like they were moving; we measure movement over the whole window instead.
Accomplishments that we're proud of
Sub-Second Incident Alerting: Successfully reducing the emergency dispatch notification pipeline from minutes down to milliseconds.
Seamless Dual-Engine Architecture: Bridging the gap between immediate reactive crisis response and long-term proactive safety intelligence in a single, unified interface.
Functional End-to-End Demo: Building a working prototype that accurately tracks pose changes, triggers live alerts on a dashboard, fires off real SMS notifications, and provides actionable AI safety coaching from simulated log data.
Real-Time on a Laptop GPU: 31 FPS on an RTX 3050, with a live fall detected at 0.88 confidence about one second after impact, backed by 107 automated tests on the vision engine.
What we learned
Closed-Loop Safety Design: True workplace safety isn't just about faster alert times; it requires coupling instant reactive emergency response with proactive data analytics to eliminate hazards at the source.
Optimizing Vision for the Edge: Balancing model accuracy with inference speed to achieve near-zero latency dispatch times on resource-constrained video feeds.
Human-Centered AI: Designing conversational AI tools that turn complex telemetry logs into actionable, easy-to-digest safety drills for busy floor managers.
What's next for Hazard Lens
Hardware Edge Deployment: Packaging the computer vision pipeline to run on local edge devices (e.g., NVIDIA Jetson) for zero-cloud latency processing.
Direct AGV/Forklift Interlocking: Transmitting real-time deceleration and emergency braking signals directly to Autonomous Guided Vehicles (AGVs) upon detecting a human hazard in transit zones.
Multimodal Audio Alerts: Incorporating text-to-speech sirens and intercom voice warnings across facility floor speakers when critical incidents occur.
Seizure Detection: flagging rhythmic limb movement after a fall.
Thermal Cameras: taking over in smoke-filled areas where normal cameras report impaired vision.
Log in or sign up for Devpost to join the conversation.