The Fifth Elephant 2020 edition

On data governance, engineering for data privacy and data science

Automatic Collision Notification (ACN): A Smartphone Based Crash Detection Technology

Submitted by Arnab Chakraborty (@arnabchak) on Apr 9, 2020

Status: Submitted

Abstract

Due to the ever increasing number of vehicles, a significant rise in road crash related loss, fatalities and disabilities has been observed in recent years. An accident detection technology can be seen as a solution to enhancing road safety by taking actionable measures,e.g. roadside assistance, notifying emergency service providers etc. Most of the existing crash detection technologies either use vehicular sensors and telematics devices attached to it – which are mostly available in more recent and expensive cars – or in case of smartphone based solutions, require access to microphone and camera – which needs extra permission for the application to execute properly. The Automatic Collision Notification (ACN) by Zendrive is a machine learning based solution to road crash detection and notification generation that leverages more than 50 billion miles of driving data in Zendrive’s possession and runs on smartphones using only the inertial, GPS and barometer sensors (optional). In this talk, we discuss how the ACN technology has been built from scratch – the challenges and our approach to solutions regarding data collection, analysis, modeling, deployment and post-production management.

Who should attend?
Anyone who is interested to see how one of the most challenging real life problems in road safety is being solved using billions of miles of driving data by Zendrive.

Outline

Introduction:
• Motivation behind driving behavior modeling based on smartphones
• Motivation behind building vehicular accident detection technology
• Ingredients required to build such a technology
Data:
• Data collection
• Label generation
• Data description
Algorithm:
• Algorithm challenges in crash detection
• Why ML-based solutions?
• Algorithm details: a fusion of DL, GBM and heuristics-based rules
• Trade-offs: Latency vs Accuracy; Precision vs Recall
Deployment:
• Challenges in deploying on smartphones
• Processing time vs callback time
• Post-production management:
• Feedback collection
• Removing labeling bias
• Performance analysis
Current performance:
• Latency specific performance

Requirements

Nothing specific.

Speaker bio

Arnab Chakraborty is currently working as a Data Scientist at Zendrive, Bangalore from 2019. Arnab has completed his bachelors (B.Stat) and masters (M.Stat) in Statistics from ISI, Kolkata. Having a PhD in Statistics with a dissertation focused on Spatial Statistics and Big Data Analytics, Arnab has more than 5 years of experience working on multiple data science problems across various domains both in industry and academia. In his pastime, Arnab likes to watch and play football and talk about almost all kinds of sports.

Links

Comments

{{ errorMsg }}

You need to be a participant to comment.

Login to leave a comment