Research Projects

The transportation section is one of the largest consumers of petroleum products, contributing to 16% of the world’s greenhouse gas emissions. Along with electric vehicles and other relevant technological solutions developed for sustainable transportation, there is an imminent need to involve humans in the loop for reduction of greenhouse gas emissions during driving. Unfortunately, human drivers have limited knowledge of how they can adapt their driving style to reduce fuel consumption, thereby leading to cost savings and reduced emissions. Eco-driving technology offers a cost-effective, scalable strategy for sustainable transportation by including humans in the loop for improving fuel efficiency. This involves providing personalised feedback to drivers on how they can adjust their driving behaviour or take necessary driving actions, which can lead to a reduction in fuel consumption and emission levels. Our project proposes development of a prototype of an Eco-Driving Assistance System (EDAS) for Indian Highways using Reinforcement Learning. Our project consists of five key objectives to achieve our desired goal. First, we propose to use dashboard cameras to estimate the surrounding traffic state condition. Drivers navigate their vehicles by acceleration/deceleration and steering angle change. Their response is primarily driven by the surrounding traffic state condition, which includes surrounding vehicles’ position and speed. In this project, we will develop computer vision models which can handle the unstructured Indian driving conditions and estimate surrounding traffic state, which goes beyond the traditional leader-follower car-following behavior. Second, we propose to use the On-Board Diagnostics (OBD) along with IMU-GPS sensor data to classify and describe (in natural language) the driving maneuver of the vehicle. This will involve developing transformer-based text generation model which can take the time series data as input to generate description of driving action in natural language (NL) for easier human understanding. Our third objective will involve developing vehicle-specific fuel consumption model fusing the data obtained from dashcam with OBD, IMU-GPS sensor data. The fuel consumption model will help to gain insights into the fuel usage behavior of the drivers. The dash-cam based surrounding state estimation, ego-vehicle driving maneuver classification, and fuel consumption model will act as input for our primary objective of this project, reinforcement learning (RL) based eco-driving strategy development. Human drivers primarily prioritize safety and urgency while choosing vehicle action. Therefore, human driving cannot be taken as a benchmark for eco-driving strategy development. This project, therefore, proposes RL-based EDAS. Our key innovations will include (i) combining longitudinal action (acceleration/deceleration) with lateral action (steering wheel angle) and (ii) incorporating multi-vehicle anticipation, to handle two-dimensional disorderly movement of Indian traffic in our RL model. Finally, we propose to develop a prototype which provides personalised feedback to drivers regarding eco-driving strategy. This will involve providing real-time feedback regarding eco-driving speed and acceleration/deceleration by comparing the human driving action with RL-based proposed action. Further, personalised feedback will be provided to drivers in natural language on how they can adapt their driving styles to improve fuel efficiency. Our proposed mobile application will continuously monitor driver action and provide feedback in NL at multiple time scales (real-time and trip-based). Further, the proposed solution can provide general guidelines to Indian drivers on how they can adapt their driving strategy for improved fuel economy, thereby allowing us to achieve our desired goal, involving human drivers in loop for developing a sustainable driving environment, reducing fuel consumption and emission levels.