VisionGuard AI

AI-powered warehouse safety intelligence for detecting, tracking, analyzing and explaining forklift–pedestrian risk.

Python Ultralytics YOLO OpenCV ByteTrack TTC Risk Prediction Advanced Prototype
View on GitHub
VisionGuard hero preview VisionGuard dashboard VisionGuard live monitoring VisionGuard event investigation VisionGuard analytics VisionGuard safety zone configuration VisionGuard core pipeline architecture VisionGuard explainable near-miss replay

Overview

Most YOLO demos stop at detection and a bounding box. VisionGuard goes further — it transforms ordinary warehouse CCTV video into explainable safety events by chaining detection, multi-object tracking, ground-plane calibration, trajectory analysis, time-to-collision (TTC) prediction and a transparent, multi-signal risk score into one pipeline.

A detection alone is not an incident — an incident requires context. VisionGuard supplies that context through zones, real-world distance, motion history, closing speed and exposure duration, then automatically saves a video clip, snapshot and JSON explanation for every critical event.

Purpose — Why VisionGuard Exists

Forklift–pedestrian near-misses are one of the most common warehouse safety incidents, and manual review of hundreds of camera hours can't scale and rarely catches risk before it becomes an accident. VisionGuard exists to close that gap: it watches every camera continuously, understands warehouse geometry, and flags a dangerous moment before or as it happens — with evidence a safety officer can actually act on, not just a raw alarm.

Animated Pipeline

1
Video Input
Warehouse CCTV / camera feed ingested frame by frame.
2
Detection
YOLO detects people and forklifts in each frame.
3
Tracking
ByteTrack assigns persistent IDs across frames.
4
Real-World Mapping
Homography converts pixels to ground-plane meters.
5
Risk Engine
TTC, closing speed, zones and exposure feed a weighted score.
6
Event Evidence
Critical events save a clip, snapshot and JSON explanation.

Key Features

Virtual Safety Zones — polygon-defined walkways, forklift lanes and restricted/danger areas per camera.
Real-World Distance — pixel positions converted to meters via ground-plane homography calibration.
Trajectory & Speed — per-track motion history drives velocity, direction and closing-speed estimates.
TTC / Closest-Approach — constant-velocity heuristic predicts whether two objects will breach a safety radius.
Explainable Risk Scoring — a weighted 0–100 score combines distance, TTC, closing speed, shared zone and exposure duration.
Automatic Event Evidence — rolling pre/post buffer saves an MP4 clip, snapshot and JSON metadata per critical event.

Benefits

Fewer Accidents

Flags dangerous forklift–pedestrian proximity before it becomes a collision.

Objective Evidence

Every event ships with a clip, snapshot and JSON — no more he-said/she-said reviews.

Scales Across Cameras

One pipeline monitors many cameras continuously without extra staff.

Actionable Analytics

Turns raw footage into hotspot and near-miss trend data for safety teams.

Risk Model Weighting

Distance30%
TTC25%
Closing Speed15%
Shared Danger Zone15%
Exposure Duration10%
Detection Confidence5%

Severity bands: 0–34 Low · 35–59 Medium · 60–79 High · 80–100 Critical.

Technology Stack

Python 3.10+ Ultralytics YOLO ByteTrack OpenCV NumPy OpenCV Homography Streamlit (planned) FastAPI (planned) Docker (planned)
Explore the Repository