How Iris Technologies Built an AI-Powered Transport Detection System That Works Across Continents

The Challenge

Iris Technologies AB, a Stockholm-based impact tech startup, set out to help businesses and individuals measure and reduce their climate impact in support of the UN 2030 Sustainability Agenda. Their vision was ambitious: create a platform that could automatically track how people travel and calculate the carbon footprint of each journey — without requiring any manual input from users.

Why This Was Hard

Unlike controlled environments where you might track delivery vehicles with dedicated hardware, this system needed to work:

• Across dramatically different traffic conditions (smooth European highways vs. congested Asian city streets)
• In areas with no internet connectivity (subway tunnels)
• Without destroying phone battery life (users won't keep an app that drains their battery)
• With high accuracy across diverse real-world scenarios (traffic jams, stop-and-go driving, multi-modal trips)
• On both iOS and Android (each with different background processing limitations)
• Within a startup budget (no room for expensive cloud computing or extensive field testing)

Traditional GPS-based solutions couldn't solve this. A car stuck in traffic looks like a bicycle on GPS. A subway train underground has no GPS signal at all. The solution required something more sophisticated: machine learning combined with sensor fusion and smart algorithms.

The Solution

Phase 1: Understanding the Problem Space

Before writing a single line of code, we conducted extensive research into existing academic papers on transportation mode detection. We analyzed the strengths and limitations of various sensor combinations and machine learning approaches.

Phase 2: Custom Deep Learning Model Development

We developed a custom Multi-Layer Perceptron (MLP) deep learning model trained on sensor data fusion from:

• Gyroscope — Detects rotational movement patterns (different for bicycles vs. cars)
• Accelerometer — Measures acceleration forces and vibration patterns unique to each mode
• GPS/Location — Provides speed and route information
• Magnetometer — Tracks directional changes and turning patterns
• Barometer — Detects altitude changes (critical for subway detection)

The Training Data Challenge

We couldn't buy a ready-made dataset for this specific use case. Instead, we:

1. Recruited students and volunteers to collect real-world data across different transportation modes
2. Supplemented with publicly available transportation research datasets
3. Built a feedback loop where users could correct misidentifications, continuously improving the model

The Geographic Diversity Problem

Early testing revealed a critical challenge: transportation patterns vary dramatically by region.

We needed the model to work in both environments without being retrained for each region. This required extensive testing and algorithm refinement to handle diverse global transportation patterns.

Technical Architecture

A full-stack solution combining native mobile apps, cloud infrastructure and machine learning.

Challenges Overcome

Challenge 1: Battery Optimization

The Problem: Continuous sensor monitoring could drain a phone battery in hours. Users would immediately uninstall an app that killed their battery life.

Our Solution:

• Implemented adaptive sampling rates (collect data more frequently when movement is detected, less when stationary)
• Used geofencing to detect when users left home/work and only then activated intensive monitoring
• Optimized sensor wake-up patterns to minimize CPU usage
• Carefully managed iOS background execution limits (more restrictive than Android)
• Reduced battery impact from approximately 20% daily drain to less than 5%

Challenge 2: Trip Start/Stop Detection

The Problem: How do you distinguish between a traffic light stop and the end of a trip? Or between a parking lot search and continued driving?

Our Solution:

• Built a time-and-distance algorithm analyzing duration, location history and time of day
• Implemented a "trip confidence score" to finalize trips only after meeting thresholds
• Handled edge cases like drive-throughs and gas station stops

Result: Accurate trip boundaries even in congested urban environments with frequent stops.

Challenge 3: Subway Detection Without Internet

The Problem: Subway trains travel underground where GPS signals are unavailable and cellular data doesn't work. How do you track a trip you can't see?

Our Solution:

• Used barometer data to detect underground descent
• Combined accelerometer patterns (distinctive subway vibrations) with last known location
• Calculated likely routes using entry/exit stations, time elapsed and known subway speeds
• Applied confidence scores and allowed user corrections

Challenge 4: Airplane Detection

The Problem: When users fly, their phones are typically in airplane mode — no GPS, no cellular data and limited sensor activity. Yet flights represent significant carbon footprints that need to be tracked.

Our Solution:

• Detected airplane mode activation during typical travel hours
• Used the last known location before airplane mode and first location after
• Calculated distance and verified it matched airport-to-airport travel
• Applied time-distance heuristics (e.g. 500+ miles in 2 hours = likely a flight)
• Prompted users to confirm flight travel rather than assuming

Result: Captured air travel carbon footprint without requiring manual entry.

Challenge 5: Multi-Modal Trip Segmentation

The Problem: In European cities especially, a single "trip" might involve walking to a bus stop, taking a bus, transferring to a subway, then walking to a destination. The app needed to recognise this as one journey with multiple segments, not five separate trips.

Our Solution:

• Developed trip chaining algorithms analyzing time gaps, geographic connections and travel direction
• Created configurable time thresholds for trip continuation vs. new trip detection
• Handled different cultural patterns (Europe: many transfers, Asia: fewer but more stops)
• Allowed users to merge/split trips if the algorithm misclassified

Result: Intelligent trip representation that matched users' mental models of their journeys.

Real-World Testing

To validate the system's robustness, we conducted extensive testing in two dramatically different environments:

Testing the model across these diverse conditions was critical. We didn't want a solution that only worked in Swedish suburbs. The 88% accuracy achieved across both regions validated that our sensor fusion and algorithmic approach was genuinely robust.

Business Value

For Iris Technologies

For End Users

Key Achievements

Lessons Learned

1. Real-World ML Is Messy

Academic papers often show 95%+ accuracy in controlled conditions. Real-world accuracy of 88% across diverse, uncontrolled environments is actually exceptional — because the real world is full of edge cases, sensor noise and unpredictable user behaviour.

2. Geographic Diversity Matters

A model trained only on Swedish data would have failed in Lahore. Testing across continents wasn't just nice-to-have — it was essential for building a truly robust system.

3. Battery Life Is Non-Negotiable

No matter how accurate your AI, users will uninstall an app that kills their battery. Mobile ML requires obsessive optimisation of sensor usage, processing frequency and background execution.

4. Sensor Fusion Beats Single-Source Data

GPS alone couldn't solve this problem. Accelerometer alone couldn't solve it. The magic happened when we intelligently combined multiple sensors and applied contextual algorithms.

5. Startups Need Technical Partners Who Understand Constraints

Iris didn't have unlimited budget or time. We delivered an MVP in 6 months with a lean team by making smart technology choices, optimising cloud costs from day one, focusing on core functionality first, and bringing deep expertise so we didn't waste time on dead ends.