Soon, vehicles equipped with Cooperative Vehicle Intelligent Transport Systems (C-ITS) will be on Australian roads. C-ITS will enable several safety functions (use cases) which are expected to improve road safety. After-market products C-ITS based has the advantage of being a feasible solution for existing vehicles with outdated technologies, which might be the origin of the road safety problems. Moreover, C-ITS data enables real-time evaluation of driver behaviour which can be used in several applications such as driver assistance and early warning of predicted dangerous situations.
Driver style identification and scoring is an active research area that is needed in vehicle insurance and fleet management industries. The models used in these applications require an accurate extraction of many figures of merit, such as lane changing and cornering.
This research investigates the adoption of a semi-supervised learning concept called co-training to train designed deep neural networks to detect aggressive lane change and cornering using the information collected from the Cooperative Awareness Message (CAM) in the connected vehicle environment.
The investigated approach will take advantage of the Ipswich Connected Vehicle Pilot (ICVP) different data modalities (i.e. the position and kinematics) to build two different deep neural networks (i.e. classifiers), one on position and the second on the other data modalities (i.e. acceleration, jerk, yaw, speed). The co-training paradigm allows using each classifier to help train the other. The developed framework will be used to automatically detect the aggressive lane changing and cornering in the ICVP test data and can be used with any similar data.
- Manually label a small subset of ICVP data. The manually labelled data will consist of two datasets for the aggressive lane change and cornering.
- Construct a visualisation (image) for each data example in the two datasets.
- Fine tune a pre-trained convolutional neural network (CNN) using the images (i.e. the first learner)
- Fine-tune one of our pre-trained Long Short Term Memory (LSTM) or Gated Recurrent Unit (GRU) models using the vehicle kinematics (i.e. second learner).
- Co-train the two learners (i.e. CNN and LSTM or the CNN and GRU).
- Test the final models.
- Write the results in the journal format.
This project aims to:
- develop an accurate deep learning-based aggressive lane change detector
- develop an accurate deep learning-based cornering detector
- write a journal paper about the developed models.
Skills and experience
You should have a mathematical background, and basic knowledge in:
- deep learning
- Tensorflow or Pytorch.
Contact the supervisor for more information.