Every year over 1.3M people die on roads. In recent years the rates of fatality and collisions have increasingly gone upward, reversing a several decade long downward trend.

At Zendrive, we use smartphone data to understand and decode unsafe driving behaviours like aggression, non- adherence to the rules of the road and distraction. Using sophisticated machine learning techniques and massive amounts of data (150B miles of data over 50M users), we have built the world’s leading driving behaviour analysis platform that has already helped save hundreds of lives.

A core component of this platform is an algorithm to detect vehicular collisions. In this talk, we aim to take you on the fascinating journey of building this algorithm through myriads of challenges - smartphone sensors, data acquisition, detection of rare events, testing, and so on. The talk will highlight how these challenges were overcome through a combination of creative problem-solving and sophisticated ML techniques.

Outline

Why have a collision detection algorithm?
- Saving lives by speeding up emergency response
- Measure of risk on the road

High level challenges
- Rare event (1 per million miles)
- Mix of time-scales
- Smartphones - no custom hardware

Building an MVP
- Where is the data?
– OEMs, Being creative with misuse!
- Algorithm
– Handling 3 phases
– Ensemble of ensembles

Curious cases
- Roller coasters, bumping into bins, skydiving!
-Continuous improvements

Data - More and more and more
- Customer feedback
- Manual review - label noise

Algorithm - sophistication
- Physics + Data + Machine learning

Standing today
- Largest repository of collision data
- Most widely used smartphone based algorithm

Looking ahead
- Deep learning

Looking back - what have we learnt?

[The sequence appears in the talk attached below (timestamp from 6 min to 27 mins)]

Speaker bio

Aditya Karnik is Director of Data Science at Zendrive. He has 14+ years of experience in academic and industrial research labs. His interests are in Mathematical modeling, Optimization & control and Predictive modeling.

Slides

https://drive.google.com/file/d/1tMWf_fa3Oe5_DOFTWa07lh_iV2oaOgJo/view?usp=sharing