Open-Architecture ADAS Research / Software v1.0 Validated

SAFAR

Affordable road-safety intelligence for existing vehicles.

A modular, sensor-agnostic framework that isolates visual object perception from deterministic spatial relevance, temporal hazard confirmation, and risk-calibrated driver intervention.

01 DETECT YOLO Vision
02 UNDERSTAND Ego Corridor
03 ASSESS Risk Engine
04 ACT Safe Intervention
Explore Technical Stack Live Interactive HUD
GENESIS & PHILOSOPHY

Why we started SAFAR

Modern high-end vehicles increasingly package advanced driver-assistance systems. But bringing comparable safety intelligence to hundreds of millions of on-road vehicles has historically required prohibitive hardware budgets or whole-vehicle replacement.

“

Can meaningful road-hazard awareness be built as an affordable, modular framework for vehicles that don't have factory ADAS?

What started as an ideathon concept quickly evolved into a deeper systems engineering challenge: detecting an obstacle is only the beginning. A real-world vehicle safety system must understand whether an object is in the projected ego corridor, whether it is closing in, whether the hazard persists across multiple camera frames, and whether driver intervention is strictly justified.

Rather than building a brittle end-to-end black box, SAFAR establishes a decoupled architecture where perception, tracking, ego-corridor geometry, multi-frame confirmation, and risk evaluation operate as distinct, testable engineering layers.

Evolution Pathway Concept to Simulation Adapter
  1. Initial Concept

    Inquire whether consumer cameras could provide affordable collision warning.

  2. YOLO 2D Detection

    Implement real-time visual classification of cars, trucks, two-wheelers, and pedestrians.

  3. "Detected ≠ Dangerous" Realization

    Discovered that raw detections trigger non-stop false alarms on parked roadside vehicles.

  4. Spatial-Temporal Tracking

    Added IoU object tracking and history buffers to maintain persistent identities across frames.

  5. Ego-Corridor & Motion Model

    Engineered resolution-independent trapezoidal corridor filtering and qualitative approach analysis.

  6. Multi-Frame Temporal Confirmation

    Designed a 5-state hysteresis machine (Candidate → Confirmed → Hazard) to eliminate single-frame noise.

  7. Closed-Loop Game Adapter

    Completed The Crew 2 simulation adapter with reversible DirectInput hardware scancode control.

SYSTEM DESIGN INSIGHT

The problem isn't detecting a car.

Standard computer vision treats every detected bounding box as an isolated entity. A reliable ADAS framework must extract actionable context from the visual scene.

Standard Object Detector

What YOLO Sees

Car 0.88
Car 0.94
Truck 0.81
Car 0.76

Indiscriminately flags all 4 objects. Cannot distinguish between a parked vehicle on the left and a braking truck in the driving lane.

VS
SAFAR Context Engine

What SAFAR Solves

01
Which object is in our path? Evaluates bottom center intersection within the parameterized trapezoidal ego corridor.
02
Which one is closing in? Evaluates image-space bounding box expansion rate without guessing metric km/h.
03
Which hazard persists over time? Requires multi-frame confirmation to prevent spurious single-frame false activations.
04
What response is justified? Maps confirmed threat severity to CONTINUE, CAUTION, WARN, SLOWDOWN, or EMERGENCY_BRAKE.

Existing Vehicle Base

Over 1.4 billion vehicles globally lack built-in ADAS. Replacing usable fleets is economically and ecologically unviable.

Hardware Cost Barriers

Automotive-grade LiDAR and heavy compute clusters are out of reach for consumer retrofits and commercial taxi fleets.

Alert Fatigue & Trust

False positive alerts destroy driver trust. A safety system that beeps needlessly will simply be switched off.

Complex Road Realities

Mixed multi-modal traffic, unmarked road shoulders, potholes, glare, and erratic lane changes demand resilient reasoning.

SYSTEM ARCHITECTURE

A modular framework, not just a detector.

SAFAR decouples perception from safety reasoning. Different sensor backends (phone cameras, webcams, simulated render targets, future stereo rigs) feed into a unified, deterministic risk and decision engine.

01 SENSOR INPUT Win32 Screen / Video / Camera
→
02 PERCEPTION YOLO Detector + Adapter
→
03 TRACKING IoU Tracker + History
→
04 EGO CORRIDOR Trapezoid Path Relevance
→
05 HAZARD ENGINE Lead Selector + Confirmation
→
06 RISK ENGINE Deterministic Threat Mapping
→
07 DECISION & ACT HUD Advisory / Game Controller
MODULE 01

Spatial & Ego Corridor

Maps dynamic vehicle corridors in normalized coordinates ($x \in [0, 1], y \in [0, 1]$) with parameterized horizon height, top width, and bottom lane width.

EgoPathModel(horizon_y=0.46, bottom_width=0.70)
MODULE 02

Temporal Confirmation

Prevents jitter through multi-frame state tracking. Single frames remain in CANDIDATE, holding state through transient occlusions via hysteresis.

Candidate → Confirmed → Hazard → Cleared
MODULE 03

Reversible Game Controller

Isolated from the risk engine. Injects DirectInput hardware scancodes with foreground window focus guards and automatic 3.0s override timeouts.

SendInput(SCANCODE_W_UP, SCANCODE_S_DOWN)
INTERACTIVE BENCHMARK

Live HUD & Telemetry Simulator

Test how SAFAR evaluates different road situations in real time. Switch scenarios on the left to observe how camera tracking, path relevance, confirmation state, and controller actions respond.

SELECT SCENARIO
VISUAL SENSOR FEED LIVE HUD OVERLAY
#image-2 car [LEAD HAZARD]
#image-1 truck [ADJACENT]
Capture: 30.2 FPS Latency: 37.8 ms ADAPTER RUNNING
TELEMETRY & DECISION DIAGNOSTICS
TARGET ID: #image-2 car
CONFIRMATION: CONFIRMED
CORRIDOR RELEVANCE: HIGH
APPARENT MOTION: APPROACHING
DISTANCE / TTC: UNKNOWN (Single Camera)
RISK / DECISION: HIGH → SLOWDOWN
SIM CONTROLLER: SLOWDOWN_OVERRIDE (Light Brake)
FAIL-SAFE HOTKEY: F8 (Press to Unlock)
SYSTEM VERIFICATION

What we've achieved so far

A concrete transition from conceptual vision to a tested simulation adapter with full deterministic test coverage.

Computer Vision Core

  • Ultralytics YOLO inference (~30 FPS on standard GPU/CPU)
  • Multi-object IoU tracking with historical trajectory
  • Ego-corridor lateral containment evaluation
  • Image-space qualitative expansion & motion analysis

Deterministic Safety Logic

  • Lead-hazard ranking via composite image proximity
  • 5-state temporal confirmation state machine
  • Hierarchical threat levels (Safe → Caution → Warn → Slowdown → Emergency)
  • Hysteresis state preservation over transient frame drops

Simulation & Intervention

  • Asynchronous The Crew 2 screen capture with zero backlog
  • DirectInput hardware scancode throttle & brake control
  • Foreground window watchdog + 3.0s override timeout
  • Instant manual safety release hotkey (F8)
67
Unit & Scenario Tests Passed
2
Skipped (Optional CARLA Backend)
0
Failures / Regressions
Scientific Rigor Statement: Automated unit and scenario tests validate software logic, state machines, and fail-safes. They serve as rigorous software validation, not proof of real-world physical vehicle safety.
ENGINEERING HONESTY

What broke — and what we learned

Reliable engineering is forged through edge cases and failure modes. We openly document what failed in early iterations and how the architecture was redesigned.

Observed Failure 01

Detection ≠ Hazard Relevance

Raw YOLO detections triggered alerts on cars parked on sidewalks or moving safely in opposing lanes, overwhelming the driver with noise.

→
Architecture Redesign

Ego-Corridor Trapezoid Filtering

Introduced resolution-independent trapezoidal corridor calculations that only elevate objects intersecting the vehicle's projected heading corridor.

Observed Failure 02

Single-Frame Noise Created False Alarms

A single flickered bounding box in bad lighting caused sudden state jumps, triggering abrupt deceleration warnings.

→
Architecture Redesign

5-State Temporal Confirmation Machine

Implemented Candidate → Confirmed → Hazard progression. Single-frame detections are classified as Candidate and cannot trigger braking interventions.

Observed Failure 03

Monocular Distance Estimation is Dangerous

Estimating physical metric distance and Time-To-Collision (TTC) from a single 2D camera without depth sensors produces untrustworthy results.

→
Architecture Redesign

Explicit "UNKNOWN" Metric Policy

SAFAR explicitly marks Distance = UNKNOWN and TTC = UNKNOWN in camera mode, making decisions purely on validated 2D expansion rates.

Observed Failure 04

Water-Filled Potholes Defeat 2D Vision

Muddy water in road potholes reflects the sky, appearing identical to flat road puddles to standard RGB convolutional models.

→
Next Research Direction

Stereo Disparity & Surface Normal Geometry

Road depression hazard identification requires multi-view stereo disparity geometry rather than raw 2D bounding box classifiers.

OBJECTIVE BOUNDARIES

Where SAFAR is today

We believe in honest, uncompromising transparency about our current technology readiness level.

✓

Validated Capabilities

  • Real-time image perception with YOLO11 backend
  • Multi-object tracking with persistent identity buffers
  • Ego-corridor spatial containment filtering
  • Deterministic risk, hysteresis, and decision logic
  • Non-intrusive, reversible game simulation adapter
  • 67 automated unit and scenario tests passing
⚠

Active Research Limitations

  • Metric depth and physical TTC not yet validated
  • Single-camera vision lacks true 3D spatial depth
  • Pothole depth measurement requires stereo sensor testing
  • Real-world adverse weather (heavy fog, monsoon rain) unvalidated
  • No physical automotive throttle/brake actuator control
  • Not certified for production-grade road deployment
🛡️
Research Prototype Notice: SAFAR is currently an experimental research and validation prototype, not an autonomous driving system. It is designed to assist driver awareness, not replace human responsibility or attention.
PHASED ROADMAP

How we are removing those limitations

Our disciplined, stepwise engineering trajectory from software simulation to physical fleet validation.

PHASE 01 CURRENT · COMPLETED

Software Core & Simulation

YOLO perception, ego-path modeling, multi-frame temporal confirmation, risk decision logic, and closed-loop The Crew 2 game testing.

PHASE 02 IN ACTIVE R&D

Geometric Stereo Perception

Integrate calibrated stereo camera pairs to measure genuine metric depth, closing velocities, and true Time-To-Collision without LiDAR.

PHASE 03 PLANNED

Road-Surface Intelligence

Deploy disparity-based surface geometry analysis for pothole depth measurement, speed breakers, and water-filled depression detection.

PHASE 04 PLANNED

Low-Cost Multi-Sensor Fusion

Fuse vision streams with vehicle IMU sensors, OBD-II speed telemetry, and low-cost ultrasonic/time-of-flight ranging arrays.

PHASE 05 FUTURE

Controlled Vehicle Pilot

Deploy non-intrusive dashboard HUD and audio advisory units on a closed-track test vehicle under controlled test scenarios.

PHASE 06 FUTURE

Commercial Fleet Validation

Run real-world telemetry pilot with regional cab and delivery fleets to evaluate safety metrics, driver fatigue, and passenger trust.

VALUE PROPOSITION

Why retrofit instead of replace?

NEW CAR WITH FACTORY ADAS
$$$$$$$$$$$$

Forces vehicle replacement. Highly capital intensive and completely inaccessible to everyday fleet operators and existing car owners.

VS
EXISTING VEHICLE + SAFAR
Fraction of Cost

Adds modern safety intelligence to vehicles already on the road via modular, affordable retrofit hardware and software.

“SAFAR is designed around the idea that safety intelligence should not require replacing an otherwise usable vehicle.”

GO-TO-MARKET

Start with fleets, not everyone.

A disciplined, data-first adoption strategy targeting high-utilization commercial operators where safety directly impacts operational margins.

01 SAFAR Core
→
02 Cab Fleets
→
03 Small Pilots
→
04 Safety Telemetry
→
05 Fleet Validation
→
06 SAFAR-Certified Cabs
→
07 Expansion

Why Cab Fleets First?

  • Driver Fatigue & Safety: Proactive awareness support during long, strenuous driving shifts.
  • Vehicle Uptime: Preventing collision damage keeps productive vehicles generating revenue.
  • Incident Telemetry: Objective driving risk logs and incident replay for fleet managers.
  • Passenger Trust: A visible, validated commitment to passenger safety.

The Fleet Validation Proposition

A SAFAR-equipped commercial vehicle can eventually carry a recognizable safety identity backed by empirical road telemetry.

Pilot Hypothesis: We plan to test whether visible, telemetry-backed safety technology influences rider preference and provides measurable fleet differentiation.
COMMERCIAL STRUCTURE

Simple Per-Vehicle Retrofit Economics

HARDWARE SAFAR Sensor Unit Low-cost camera & compute
+
SOFTWARE On-Device Intelligence Perception & risk engine
+
FLEET CLOUD Telemetry Dashboard Safety analytics & reporting
=
REVENUE MODEL Per-Vehicle Tier One-time retrofit + recurring SaaS
INITIAL TARGET

Cab & Taxi Fleets

Ride-hailing operators and commercial taxi associations.

EXPANSION 01

Delivery Logistics

Last-mile e-commerce delivery vans and intra-city cargo.

EXPANSION 02

Buses & Transport

Corporate shuttles, school buses, and public transport.

LONG TERM

Consumer Retrofit

Aftermarket safety modules for general car owners.

THE BUILDERS

Built by

A dedicated student engineering team building SAFAR from concept to working prototype.

SD

Saksham Dixit

Student Engineering Team

KA

Krish Agarwal

Student Engineering Team

YA

Yazdaan Ansari

Student Engineering Team

github.com/SenorDan031/SAFAR
COLLABORATE WITH US

Want to build safer roads with us?

We are actively looking for mentors, pilot partners, incubators, mobility startups, fleet operators, and researchers working on road safety.

Contact via Email View GitHub Repository
Direct Email: agarwalpk2301@gmail.com · Open for Mentorship & Fleet Pilot Trials