An AI system for automated profiling of driving patterns

Lead Research Organisation: University of Warwick
Department Name: Mathematics

Abstract

In this project PhD project, in collaboration with a dynamic start-up, Evezy, we will develop an AI system for the automated profiling of driving patterns using historical measurements generated by a large fleet of vehicles and the detection of risky behaviour in real-time. This aligns in particular with EPSRC research aims for artificial intelligence technologies, complexity science, and infrastructure and urban systems.

Vehicle fleet management businesses rely on safe drivers to support their operations. Safe drivers are preferable as they have a direct economic and financial impact on the fleet operator's business. Aggressive driving increases fuel consumption, cost of vehicle maintenance, risk of accidents, cost of insurance, and risk of poor customer experience (e.g. for taxi related operations). This work aims to assess sensor information for risk identification, in a manner that is not only accurate but has an explanatory process or assessment criteria. While something like a deep neural networks may be necessary to process and classify characteristics from a raw image data-set, the implications of the identified characteristics can be described by a myriad of plausible Reinforcement Learning techniques. Potential impacts or applications include utilising images captured by sensors on automotive vehicles to identify, classify and react to possible threats to driver and pedestrian safety.

The system will implement a metric score reflecting the degree of risk that a driver exhibits at any given time based on a combination of historical as well as current driving data. The AI system will leverage a large volume of historical data collected by Evezy customers characterising how the vehicles are being operated and in what driving conditions, at a temporal resolution of approximately 2 seconds. Measurements generated from the on-board telemetry system and other sensors (such as GPS coordinates, acceleration/breaking, front camera video feeds, battery levels, mileage, vehicle's status indicating whether the engine is operating, doors are locked, and others) will be overlaid with publicly available data (such as street maps and places of interest, historical traffic conditions and accidents, speed limits, etc.) and mobile usage data (Evezy customers have to use the Evezy app to unlock and start the car, as well as to swap vehicle). Using this unique historical database, we will implement a robust risk scoring system whereby higher scores indicate riskier driving behaviour, and a tracking system for adaptively assigning risk scores and raising alerts in real-time.
Machine learning methodologies will be used to processes the data to identify key characteristics. However, a standard black box algorithm approach may not have sufficient capacity to perform this task, the lack of descriptive power can create an impasse to being implemented for real-world purposes where risk to human safety is concerned. Therefore, while something like a deep neural network may be necessary to process and classify characteristics from a raw image data-set, the implications of the identified characteristics can be described by several possible flavours of Reinforcement Learning (RL); Deep RL, Model-Based RL, or even Hierarchical RL which creates an encoding of a kind of memory into the algorithm so as that sequences of events can be identified as potential problems not just the event at one particular instant of time. It may also be necessary to compare several different methods and optimise the trade off between computational efficiency and descriptive power.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S515681/1 01/10/2018 30/09/2022
2106499 Studentship EP/S515681/1 24/09/2018 30/09/2022 Richard Fox